DeepSeek V4.1 Flash 정식 출시: 새로운 아키텍처로 552B, MIT 가중치, 그리고 일어나지 않은 V4 Pro 서비스 종료

DeepSeek V4.1 Flash is no longer a two-day experiment. The model that spent September 8 and 9 sitting inside the vendor's live API under an ID carrying its own expiry date — deepseek-v4.1-flash-expires-on-0910 — was released for real on September 10, 2026, across DeepSeek's app, web interface and API, where it now answers to the model name deepseek-flash. Three things about the release are larger than the beta implied. The first is the paperwork: DeepSeek published a technical report — titled, without much subtlety, Pushing the Limits of KV Cache Compression — and MIT-licensed weights for a 552B-parameter multimodal mixture-of-experts model built on a new Causal Encoder-Decoder architecture. Those weights are not a promise: 48 safetensors shards, about 510 GB in total, went online the morning of release, which turns the beta's unverified “new architecture, native multimodal” slogan into something a reader can download and inspect. The second is the shape of the line-up. DeepSeek V4 Flash and DeepSeek-V4-Flash-Vision-Exp are retired, their old IDs still resolving to the new model at Flash rates, and the plan to phase DeepSeek V4 Pro out on September 14 was withdrawn before the date arrived: V4 Pro is still served, at unchanged billing, and DeepSeek has promised notice before anything changes. The cheapest model in the family has taken the flagship's place on price and on DeepSeek's own benchmarks, and DeepSeek says the architecture underneath it was designed to scale to larger models. What sits above it is, as of September 25, one tracker's guess — a three-tier Flash, Pro and Ultra line at 522B, 2T and 8T — offered by the account that posted it as speculation and confirmed by nothing DeepSeek has published. The third is the one the beta never had. Artificial Analysis has now published the first outside evaluations of the model — an Intelligence Index score of 39 in Reasoning mode at maximum effort for the September weights and 45 for the 1017 checkpoint (a six-point gain that ties it with GLM-5.3 for second among open-weights models, a point behind MiMo-V2.6-Pro), against 36 for DeepSeek V4 Pro 0813 on the same scale, which makes the 552B model the highest-scoring DeepSeek model the site has measured. Ten further items are younger than the release, and they point in different directions. On September 25 ARC Prize's own model-testing lead opened a pull request adding a first-class DeepSeek V4.1 Flash profile to the foundation's benchmarking harness — the evaluator whose board this field reads as a reasoning ceiling rather than a productivity check, arriving with no result attached and none promised. Three days later an outside lab published the opposite kind of news: a finished measurement, on a different board, that puts DeepSeek V4.1 Flash next to another open-weight model and finds it the weaker of the two — and finds the number moving more with the harness wrapped around the model than with the model itself. On Exploit Bench, a Carnegie Mellon benchmark scoring how far an agent climbs toward full exploitation of a production V8, DeepSeek V4.1 Flash averaged 4.7 of 16 under the benchmark's own default configuration and 9.3 under Claude Code, while GLM-5.3-Flash peaked at 10.6 under Codex and fell to 6.2 under the very harness that suits DeepSeek V4.1 Flash best. The third is the least dramatic and the easiest to check: on October 7 the vLLM project published its own three-week accounting of serving this model, reporting its current build 1.9× faster at low concurrency and 5.3× the throughput under a 150 TPS constraint against its own day-0 implementation — an engine's before-and-after on itself rather than another outside evaluation, and the subject of the speed section below. A fourth is about what the model costs to use rather than what it scores, and it moved twice in three days: on October 12 OpenCode, the coding agent that prices this model by monthly usage allowance rather than by metered tokens, raised its allowance for it from $60 to $80 — the platform's own statement about a product it sells, carried on its own pricing page — and on October 15 it began serving the same model free, under a second model id, on a 200,000-token context against the 1M window the metered entry carries. A fifth, from October 14, measures the model's cost to serve rather than its quality: Artificial Analysis added DeepSeek V4.1 Flash to AA-AgentPerf, its hardware benchmark for agentic inference, with results on both NVIDIA and AMD accelerators — at 100 tokens per second per user, NVIDIA's GB300 NVL72 leads the field at roughly 261 million tokens per second per megawatt, and the same runs serve the model on as few as two accelerators, which is the section below in a second currency. The sixth is the release itself, superseded: on October 17 DeepSeek replaced the checkpoint behind the same deepseek-flash model name with DeepSeek-V4.1-Flash-1017 — new weights, a vendor benchmark table that beats the September model on every row it publishes, the same rate card and the same 1M-token envelope — and, as of October 18, an outside measurement of the new weights that contradicts the row the vendor leads with: Artificial Analysis scores the 1017 checkpoint 45 on its Intelligence Index and 33.8 on Terminal-Bench 4.0 under its own harness, against the 42.3 on DeepSeek's card. The seventh item is the argument that gap produced. On October 19 a tracker account with a record of calling these releases early declared itself “dead certain” that the 42.3 is real “because it's better than GLM-5.3,” and asked Artificial Analysis to run the model in a native harness. The site had already answered the day before, with a number eight and a half points lower. The eighth is not a score: on October 20 a DeepSeek tracker posted an agent transcript in which the 1017 checkpoint declined a jailbroken persona with dsh-purge — a plugin built to strip the harness's own constraints — applied, and a third party's abliteration notes put the untouched checkpoint's refusal rate at 97 of 100 harmful prompts in chat mode and 98 in think mode, with DeepSeek publishing no safety evaluation of its own for either checkpoint. The ninth is not a score and not a price: this week DeepSeek opened the inference engine its own service runs on, with a decode curve for the 1017 checkpoint and the production figures behind it, and the thread that surfaced it this week lays that curve against a September 30 Ascend 950DT benchmark that puts Huawei's part roughly five to seven times behind per chip — the serving-silicon section below. The tenth is not a score, a price or a checkpoint: on October 22 YuyuantanTan, the current-affairs account run by China Media Group, reported that DeepSeek and MiniMax had completed the registration procedures attached to China's draft technology-export catalogue and finished open-sourcing and overseas licensing. That is state media reporting on two companies rather than a regulator's notice or a company statement, and it is the development that answers the question the draft catalogue opened on October 16 — whether the MIT weights this piece is about can keep being published and shipped abroad. The licence section below separates what the report says from what it does not.

Everything below keeps two kinds of claim apart. The specifications, benchmarks and prices come from DeepSeek's own changelog, model card and technical report — vendor-published, vendor-run, and not reproduced by anyone else. That gap has since narrowed by four independent instruments — a composite index, a community kernel board, a harness-controlled cyber benchmark and a hardware serving benchmark — one of which, the index, has now been run against both checkpoints. None of them replicates a single vendor row, and ARC Prize is still building the reasoning evaluation, with nothing published from it yet. DeepSeek replaced the model behind the same API name on October 17 with the DeepSeek-V4.1-Flash-1017 checkpoint. For a day that meant every outside number in this piece described weights the vendor no longer served, with no external score for the ones it did; on October 18 Artificial Analysis re-ran its index on the new checkpoint and published 45, against 39 for the September weights. The other three instruments have not been re-run: the cyber benchmark, the kernel board and the hardware benchmark still describe the September weights, and no independent score, old or new, replicates one of DeepSeek's own benchmark rows. The cost-of-serving figure the vLLM project published on October 17 is newer only in publication date: the chart behind it is dated October 16, the day before the swap, so it describes the same superseded weights. On its current index (v4.3.2) Artificial Analysis scores the September DeepSeek V4.1 Flash checkpoint at 39 and the 1017 checkpoint at 45, describing both as well above the median of 19 for open-weight models of comparable size; they are genuine third-party numbers and the first in this story that DeepSeek did not produce. Each is one composite of ten evaluations, not a replication of any individual vendor row, and it says nothing about the 1M-context economics. The refusal figures in this piece are a fourth kind of claim and are labelled where they appear: they come from people who build uncensored derivatives of these weights, not from DeepSeek and not from an evaluation lab, and DeepSeek has published no safety evaluation for either checkpoint. The hardware numbers in the serving section are a further class again: the NVIDIA decode curve and the production figures are DeepSeek's own measurements of its own engine, published with the code this week, and the Ascend points come from a benchmark sheet reproduced in a thread on X that is not published on either vendor's site. Where a number belongs to DeepSeek, this piece says so. One more class sits outside that taxonomy. The October 22 report that DeepSeek and MiniMax completed registration under the draft export catalogue is state media — China Media Group's YuyuantanTan, repeated across Chinese financial outlets — not a Ministry of Commerce notice, and DeepSeek has published no statement confirming it. The catalogue itself is a government draft open for comment until November 15, and is described as a draft wherever it appears here. The build names attached to the report — V4.1F-1017 and M3.1-Flash — come from a single X account on October 23 and do not appear in the state-media report or the outlets that relayed it; the checkpoint that post names had already been downloadable for six days when it was written.

10월 17일 체크포인트: DeepSeek-V4.1-Flash-1017이 바꾼 것

Five weeks after the release this piece is built around, DeepSeek replaced the model underneath it. The changelog entry dated October 17 is three sentences long: “DeepSeek V4.1 Flash has been upgraded to DeepSeek-V4.1-Flash-1017. The official web and APP models have also been updated to the new version. The API calling method and pricing remain unchanged — simply use deepseek-flash to access the latest model.” Everything in that quote is the vendor speaking about its own product, and all three claims are checkable on DeepSeek's own pages rather than taken on trust: the model-version row on the Models & Pricing table now reads DeepSeek-V4.1-Flash-1017 where it read DeepSeek-V4.1-Flash, the rate card beside it is the same off-peak and peak split published on September 10 — $0.003, $0.15 and $0.60 per 1M tokens off-peak, doubling to $0.006, $0.30 and $1.20 inside the weekday peak windows — and the API quick start still names deepseek-flash as the model string, with the 1M-token context, 384K output ceiling and 2,500-request concurrency limit unchanged. The retired alias IDs behave the way the September note said they would: deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and now answer as DeepSeek-V4.1-Flash-1017 at Flash prices. DeepSeek announced the same change from its own account that day, and the post is worth reading alongside the changelog because it is shorter and blunter: “DeepSeek-V4.1-Flash-1017 is here! Stronger benchmark performance, faster task completion, and shorter outputs than the previous version—a Pareto improvement in performance and cost. Now live on our website, app, and API,” followed by the same three harnesses and the same instruction to upgrade to DeepSeek Harness 0.2.1-rc.1 or later. Everything in it is the vendor describing its own product on its own channel: “now live on our website, app, and API” is the changelog's “the official web and APP models have also been updated” restated, and the benchmark, runtime, output-length and Pareto claims are the model card's — with one difference that matters for anyone deciding whether to try it. The card carries a number next to every benchmark row; the post carries none, and neither the post nor the changelog puts a figure on the runtime and output-length claim. Two of the four things DeepSeek is leading with here are still vendor assertions rather than published results, and the chart behind the Pareto line — the one attached to the post — is DeepSeek's own, drawn on its own task-cost measurements.

The weights moved with it, and that is the part a reader can verify without trusting a changelog. A second repository, deepseek-ai/DeepSeek-V4.1-Flash-1017, went up on Hugging Face in the early hours of October 17 under the same MIT licence: 48 safetensors shards again, the same total of about 510 GB, every shard the same size as its September counterpart, and every shard's hash different. Same shape, different contents — which is what “a minor update” should look like from the outside, and which also means the architecture claims this piece has been reporting since September are untouched: the config file that declares the 552B backbone, the 384 routed experts, the 8B-active-parameters-per-token prefill path and the 16B decode path is byte-identical to the one in the September repository, so the 890-byte-per-token KV cache is the same design rather than a new one. What the new repository does not carry is the paperwork. There is no technical report in it, no evaluation folder and no Jinja chat template; the model card says the release does not include a chat template and points instead at a Python encoding reference and a new Rust toolkit, deepseek-recipe, for turning Chat Completions, Responses and Anthropic-format requests into the model's prompt format. Anyone pinning weights by repository name should know the two downloads differ in more than their hashes. The file list says exactly how: all 48 shards carry a different checksum and an identical byte size, while the files that are not weights — the config that declares the architecture, the licence and the whole encoding folder — are byte-for-byte the same in both repositories. Seven files exist in September's repository and not in this one, and nothing was added to replace them: the technical report, the evaluation folder, the chat template, the two figures the card used to embed, and a .gitignore. The model card narrows one serving parameter as well. September's card recommended either a top_p of 0.95 or 1.0; the October 17 card recommends 1.0, with temperature at 1.0, a 1M context window and a max_tokens of at least 256K unchanged.

DeepSeek's own benchmark table is where the update is meant to show, and it is a vendor table of exactly the kind the September rows were: same-company runs, unreproduced by anyone, published on a model whose weights had been public for a few hours. Against the September checkpoint it is up on every row the card carries — HLE without tools 42.6 against 36.8, or 44.0 against 39.1 on the text-only subset; Terminal-Bench 4.0 42.3 against 31.2; DeepSWE v1.1 77.0 against 74.2; ProgramBench 29.1 against 20.3; NL2Repo-Bench 68.2 against 65.4; Automation-Bench v1.0.6 60.5 against 54.8; Agents' Last Exam 35.4 against 31.8; Apex Bench v1.1 72.1 against 64.6; Chartography with tools 83.0 against 78.9; BabyVision with tools 92.5 against 89.6; and ZeroBench-main with tools 51.0 against 49.0, at Pass@5. The same table places 1017 against the frontier models DeepSeek benchmarks against: it now clears GPT-5.6 Sol on Terminal-Bench 4.0, 42.3 against 39.9, while still trailing Opus 5 on 51.8, and it leads both on DeepSWE v1.1, 77.0 against 73.0 and 74.0 — a row read off the chart attached to the vendor's own announcement, and the single most quoted number of the week. On NL2Repo-Bench it moves to 68.2 and still sits behind Opus 5 on 75.3, which is the shape the September table had too: strongest on harnessed agentic work, behind on the rows that reward reasoning depth. The Terminal-Bench 4.0 row is the one to read twice, because it is now the only row in that table with an independent counterpart. DeepSeek ran it under the Claude Code harness; Artificial Analysis runs the same 66 tasks under mini-swe-agent, and its October 18 result for these weights is 33.8 — eight and a half points below the vendor's 42.3. On DeepSeek's own table that row clears GLM-5.3 by 4.4 points, 42.3 against 37.9. On Artificial Analysis's harness the same weights trail GLM-5.3 by eight, 33.8 against 41.9.

One line of small print is worth more than several of those deltas. The September card evaluated Terminal-Bench 4.0 with DeepSeek Harness in its Minimal mode; the 1017 card says it used the Claude Code harness for Terminal-Bench 4.0 and Mini-SWE-V2 for DeepSWE v1.1. The harness study further down this piece measured seven points of spread on one model across seven agent CLIs, so an eleven-point Terminal-Bench 4.0 move that arrives with a different harness under it cannot be read as eleven points of model. The card states which harness it used, which is what makes the row auditable rather than merely impressive, but the row is two changes wearing one number. The independent run separates them, and not in the vendor's favour. Artificial Analysis evaluated the same 1017 weights on Terminal-Bench 4.0 under mini-swe-agent — the harness it holds constant across every model on that board — and got 33.8, against 26.8 for the September checkpoint on the same harness. On a fixed harness the checkpoint is worth seven points, not eleven, and DeepSeek's headline number sits eight and a half points above what the same weights score when the harness is held constant across models. The rest of DeepSeek's framing is about cost rather than score, and it is the same axis the September release was sold on: the changelog claims the update reduces task runtime and output length, calls the result “a new performance–cost Pareto optimum across multiple harnesses,” and says it optimised for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy, recommending DeepSeek Harness 0.2.1-rc.1 or later. Runtime and output length are precisely the quantities the verbosity evidence in this piece tracks from the outside, and DeepSeek now claims to have cut both — with no figure attached. One outside measurement of the new checkpoint now exists, and it does not show the runtime or the output cut: Artificial Analysis's index run has the 1017 weights at 288.8 seconds per task against 289.9 for September, and 89,495 output tokens per task against 88,574 — flat on both, with the cost saving landing somewhere else entirely. The post DeepSeek published on its own account the same day compresses all of that to “faster task completion, and shorter outputs,” then calls the pair a Pareto improvement in performance and cost — the changelog's claim in shorter form, from the same source, and just as unmeasured. It is the vendor restating itself, not a second source — which matters, because the number off that chart is the one the outside world has spent the week repeating. The one artefact in the announcement that does carry numbers is the chart attached to the post, and it is DeepSWE v1.1 Pass@1 plotted against average cost per task under a mini_swe_v2 harness. Its caption is worth reading before the Pareto line is: the public leaderboard data underneath it is dated September 22 on the chart itself, three and a half weeks before the checkpoint the chart is being used to promote, and the comparison lines for OpenAI's models are credited to OpenAI's own DeepSWE chart rather than to a shared run. One point on that plot is new. Everything it is measured against was published before it existed.

모델 교체와 같은 날, 그 표에 대해 가장 많이 공유된 해석은 모델이 어디서 실행되는지에 이해관계가 있는 사람에게서 나왔다. Databricks에서 AI 시스템을 연구하는 Yuchen Jin은 DeepSeek-V4.1-Flash-1017이 “DeepSWE v1.1에서 77점을 기록한다”고, “83배 낮은 비용으로 Opus 5급 성능”이라고, 그리고 “우리가 이를 Databricks에서 출시할 예정”이라고 게시했다. 77은 공급사 발표에 첨부된 차트에서 가져온 DeepSeek 자체 행이며, 회사 외부에서는 아무도 이를 재현하지 못했다. 두 비교 포인트는 Jin의 것이며, 외부 확인을 견디는 것과 그렇지 않은 것으로 깔끔하게 나뉜다. Opus 5급이라는 점은 유지된다. DeepSWE v1.1을 공개하는 독립 리더보드들은 자체 실행에서 이 모델의 이름을 그 부근에 놓는다. 91개 저장소에 걸친 113개의 원본 과제를 mini-swe-agent 하네스로 채점하는 Mercor의 APEX 리더보드는 Opus 5 Max를 71.4%, DeepSeek V4.1 Flash Max 항목을 71.7%로 올리며, 같은 벤치마크의 llm-stats 보드는 Claude Opus 5를 68.8%로, DeepSeek-V4.1-Flash를 74.2%로 놓는다. 이 쌍에 대한 정직한 해석은 게시물이 만드는 것보다 좁다. 이 벤치마크를 독립적으로 측정하는 두 리더보드는 이 모델이 적어도 해당 벤치마크에서 Opus 5와 동등하다는 데 동의하며, 어느 보드도 77을 보여주지 않는다. 둘 다 9월 수치나 자체 실행 결과를 싣고 있고, 어느 쪽도 1017 체크포인트로 항목을 표시하지 않기 때문이다. Opus 5급이라는 것은 어떤 범위에 관한 진술로서 뒷받침 가능하다. 그것을 만들어낸 숫자는 아직 재현 가능하지 않다.

글의 83×에 해당하는 절반은 발표된 수치가 아니라 그의 산술이며, 바로 그 절반을 느슨하게 잡아야 한다. 어떤 리더보드도 1017 체크포인트에 대한 작업당 비용을 담고 있지 않고, DeepSeek 자체의 파레토 주장 뒤에 있는 공급업체 차트는 작업당 비용을 10센트에서 20달러까지 이어지는 로그 축에 그린다 — 이는 그렇게 큰 배수가 요구하는 종류의 격차다. 정가만 기준으로 보면 격차는 83×보다 좁다: Claude Opus 5는 주중 피크 시간대에 100만 입력 토큰당 $5, 100만 출력 토큰당 $25를 청구하는 반면 DeepSeek-V4.1-Flash는 $0.30과 $1.20이므로, 입력에서 약 17×, 출력에서 약 21×이며, 답은 캐시 적중 비율과 작업이 소비하는 토큰 수에 따라 크게 달라진다. 비용 논증의 방향은 여기서도, 이 글의 어디에서도 이견이 없다. 구체적인 배수는 아무도 공개하지 않은 비교에 관한 주장이다.

Databricks는 그 게시물에서 독자가 문서와 대조해 확인할 수 있는 부분이며, 문서는 게시물보다 더 많은 내용을 말한다. DeepSeek V4.1 Flash는 DeepSeek 출시 다음 날인 9월 11일부터 Databricks Unity Gateway에 있었고, 서비스 이름 system.ai.deepseek-v4-1-flash로 플랫폼의 Foundation Model APIs를 통해 텍스트와 이미지 입력을 지원한다. 10월 7일에는 플랫폼이 이 모델에 대해 우선순위 종량제(pay-per-token)와 예약 프로비저닝 처리량(provisioned throughput)을 추가했다. 이 모델은 또한 이 제품군에 대해 플랫폼이 지정한 대상이기도 하다. Databricks는 DeepSeek V4 Pro (0813)의 서비스 종료를 10월 30일로, DeepSeek V4 Flash (0731)의 서비스 종료를 11월 5일로 예정하면서 두 안내문 모두에서 DeepSeek V4.1 Flash를 권장 대체 모델로 명시했다. 하지만 Databricks의 어떤 페이지에도 10월 17일 체크포인트가 나온다는 말은 없다. 그 주장은 한 직원이 개인 계정으로 한 말로, 출시된 날짜가 아니라 의도에 관한 내부 신호이며, 지원 모델 페이지에는 여전히 체크포인트 라벨 없이 해당 모델이 실려 있다. 따라서 이번 주의 가용성 변화는 예정된 것으로, 플랫폼 자체가 아니라 플랫폼 내부 소식통에서 나온 것이다. Databricks 고객이 오늘 호출할 수 있는 것은 9월 가중치이며, 이 글에서 계속 설명해 온 교체, 즉 같은 이름과 같은 요금표 아래에 새 가중치가 들어가는 방식이 그 대체가 실제로 이루어질 때의 모습이다.

So what did not change is as important as what did. Three of the four independent instruments still cover the September weights and only those: the Exploit Bench harness table, the community kernel board and the AA-AgentPerf hardware runs were all produced before the swap, and the two DeepSWE v1.1 boards that carry this model's name publish an entry with no checkpoint on it. The fourth changed on October 18: Artificial Analysis added a separate row for the 1017 checkpoint, scored it 45 on its Intelligence Index and 33.8 on Terminal-Bench 4.0, and left the September row standing at 39 and 26.8. That is the first time a number in this piece was produced by someone other than DeepSeek against the weights DeepSeek is currently serving. The API change is a checkpoint refresh behind a name DeepSeek had already documented as pointing at the latest Flash model — the opposite of the September plan to make deepseek-v4-pro answer as a different model, which is the distinction developers were drawing when they objected to that one. For a caller, the practical effect is narrow and concrete: asking for deepseek-flash now gets different weights at the same price, with a better vendor-reported table behind them and, since October 18, one third party scoring them — lower on the one row where both exist. On OrcaRouter the model id and the rate card are unchanged too — DeepSeek V4.1 Flash is routed at DeepSeek's own list price with 0% markup passed through — which is what a vendor refreshing a model without moving its name or its price should look like from the caller's side of the line.

9월 10일에 실제로 출시된 것

The changelog entry opens with a single sentence: "Today, we officially release the DeepSeek-V4.1-Flash model." DeepSeek describes it as the smallest member of a new model-structure family, with native visual understanding, and it is the first release in this cycle to arrive with a technical report attached. The API side is a name swap rather than a migration:

• Model name — deepseek-flash on DeepSeek's own API; the base URL, the key and the request format are unchanged.

• 서비스 종료 — DeepSeek V4 Flash와 DeepSeek-V4-Flash-Vision-Exp는 오프라인 상태입니다. 기존 ID는 계속 허용되며 이제 V4.1 Flash로 라우팅되고 Flash 요금으로 청구되므로, 기존 통합은 코드 변경 없이 계속 작동합니다.

• V4 Pro — DeepSeek announced that from 12:00 Beijing time on September 14, deepseek-v4-pro requests would route to V4.1 Flash at V4.1 Flash prices until a V4.1 Pro shipped, then reversed the decision before the date arrived. V4 Pro is still served, at unchanged billing, and DeepSeek says it will give notice of any further change.

• Envelope — 100만 토큰 컨텍스트 윈도우, 최대 출력 384K, 계정당 동시 요청 2,500개, 이미지 입력 지원, 사고 모드 기본 활성화 및 비사고 옵션 제공.

• 가중치 — 같은 날, 저장소 deepseek-ai/DeepSeek-V4.1-Flash가 MIT 라이선스로 Hugging Face에 올라왔습니다: 총 약 510GB에 달하는 48개의 safetensors 샤드, 그 옆에 PDF로 제공된 기술 보고서, 그리고 inference 폴더에 담긴 참조 추론 구현.

That last bullet is also where the retired model ID story lands for anyone with deepseek-v4-flash hardcoded in a production path: it still resolves, but what answers is a different model. DeepSeek has been explicit that this is a compatibility shim rather than a promise, which makes it a migration deadline wearing a deprecation notice.

아키텍처, 이제 공식 기록으로 남다

이틀 전만 해도 이 섹션은 아무도 발표하지 않은 것들의 목록일 수밖에 없었다. 기술 보고서가 이를 확정했고, 그 구조적 주장은 마케팅이 아니라 문자 그대로였음이 드러났다. DeepSeek V4.1 Flash는 552B 파라미터 멀티모달 전문가 혼합 모델이며, 새로운 부분들은 바로 실행 비용을 결정하는 부분들이다:

• 백본 — 총 552B 파라미터, 레이어당 1개의 공유 전문가와 384개의 라우팅 전문가로 구성된 전문가 혼합(mixture-of-experts) 구조이며, 토큰당 6개의 라우팅 전문가가 활성화됩니다.

• CED — 40개 층의 트랜스포머로, 20개 층의 인과적 인코더와 20개 층의 디코더로 분할됩니다. 디코더의 전역 KV 캐시는 각 디코더 층 자체의 은닉 상태에서 파생되는 대신 인코더의 최종 은닉 상태로부터 투영되며, 대부분의 메모리 절감이 바로 여기서 발생합니다.

• 활성화 비대칭 — prefill 중에는 토큰당 약 8B개의 매개변수가 활성화되고 decode 중에는 16B개가 활성화되므로, 긴 프롬프트를 읽는 것이 그로부터 생성하는 것보다 상당히 저렴합니다.

• KV 캐시 — 보고서에 따르면 전역 KV 캐시는 토큰당 890바이트로, DeepSeek V4 Flash의 약 4분의 1 수준이며, 16개 채널당 스케일 하나를 사용하는 FP4 메인 KV 캐싱과 계층 간 재사용을 활용합니다.

• 어텐션 및 디코딩 — 전체, 재인덱스 및 재사용 모드와 계층적 희소 인덱서를 갖춘 압축 희소 어텐션 2, Single-Pass mHC, 그리고 DSpark 추측 디코딩.

• Engram — 토큰 기반 조회를 통해 드물게 접근되는 196B 매개변수 조건부 메모리 테이블로, 어텐션 스택 내부가 아니라 전문가 혼합(mixture-of-experts) 경로 옆에 자리합니다. 이는 이번 릴리스에서 단일 최대 가중치 블록이며, 모델 카드에 기재된 이 수치가 다운로드 크기를 가늠할 수 있게 해줍니다.

• 비전 — 2계층 MLP 프로젝터를 갖춘 DeepSeek-ViT로, 나중에 텍스트 기반 위에 덧붙이는 방식이 아니라 언어 모델 사전 학습의 시작부터 훈련됩니다.

• 훈련 — 45T 토큰의 멀티모달 코퍼스로, 희소 어텐션은 64K 시퀀스 길이에서 훈련되고 컨텍스트는 34T 토큰 지점에서 1M으로 확장됩니다. 포스트트레이닝은 SFT, 그다음 RL, 그다음 온폴리시 증류를 실행합니다.

• 추론 강도 — 1부터 100까지 연속적으로 제어 가능한 정수로, DeepSeek V4 Pro가 제공하는 낮음, 높음, 최대 등급보다 더 세밀한 조절입니다.

KV 캐시 수치는 스펙 시트를 넘어서는 파급 효과를 지닌 항목이며, 바로 이 지점에서 그 설계가 제값을 합니다. 1M 토큰 컨텍스트 창은 이를 채울 여유가 있을 때에만 유용합니다. 이전 세대의 롱컨텍스트 경제성이 고통스러웠던 이유는 두 가지입니다. 전역 캐시가 프롬프트에 어텐션하는 모든 레이어마다 커졌고, 영구 캐시 용량의 대부분이 몇 분 안에 쓸모없어지는 슬라이딩 윈도 상태를 보관하는 데 소모되고 있었기 때문입니다. 보고서의 해법은 두 문제를 모두 정면으로 겨냥합니다. 실제 동시성 조건에서 1M 토큰을 다룰 때 서비스되는 제품이 그에 부합하는지는 여전히 제3자의 검증이 필요한 종류의 주장입니다 — 현재 존재하는 독립 점수는 모델을 측정할 뿐, 롱컨텍스트 경제성을 측정하지는 않습니다 — 하지만 이제 그 메커니즘은 추론된 것이 아니라 문서화되었으며, 아래 섹션에서 이를 해부합니다.

캐시 내부 살펴보기: 토큰당 890바이트는 실제로 어떻게 달성되는가

The report's title is the thesis. Read it and the efficiency claim stops looking like a side effect of the mixture-of-experts backbone and starts looking like the design brief — which is the reading independent write-ups have converged on since release. The reason it matters commercially rather than academically is where an agent's bill actually lands: DeepSeek's framing is that the cache-hit line is the one that dominates it on long agentic workloads, which makes the size of the cache a first-order cost lever. DeepSeek's numbers put the global cache at 890 bytes per token, roughly 437× smaller than the original DeepSeek V1 and about a quarter of DeepSeek V4 Flash. Run the arithmetic on a full window and it is the difference between roughly 890 MB of global KV for a single 1M-token sequence and about 3.5 GB at V4 Flash's per-token rate. Three changes do most of the compressing: one on how much is stored, one on what gets read, and one on what is kept at all.

The storage format is the most legible of them. The main global cache is held in FP4 using the E2M1 layout with one E4M3 scale per 16 channels — half the width of the FP8 cache DeepSeek V4 stored and a quarter of FP16. What makes that survivable is that DeepSeek quantized the cache during training rather than at serving time, so the noise the narrower format introduces is something the model learned to work around instead of something inflicted on it afterwards. Not everything was compressed that way: the local sliding-window cache stays in FP8, because the report treats it as the part that is sensitive to further narrowing.

Compressed Sparse Attention 2 handles what gets read. Each attention layer is statically assigned one of three modes — Full, Reindex or Reuse — so decoder layers can share key state and indexer state instead of each keeping its own, and the first Full-mode layer builds a candidate pool that later indexing layers are restricted to. Report figures put that pool at 16,384 positions with the sparse selection over the top 512 entries, which is what decouples indexing cost from context length: DeepSeek's claim is that moving from 4K to 1M context adds only around a quarter more compute per generated token, rather than the roughly window-proportional growth the earlier design implied. That single figure is the most useful thing in the report for anyone trying to price a long-context workload, and it is also the one with no outside confirmation at all — none of the outside measurements in this piece has tested the far end of the context window.

The change doing the most work on the persistent side has the least promising name. DeepSeek V4 kept two things in its persistent cache for more than 72 hours under typical workloads: the global KV and the sliding-window attention state. Sliding-window state matters for minutes, but it was consuming close to half of the persistent cache capacity — storage sized for days spent on data with a minute of usefulness. V4.1 Flash drops SWA state out of the persistent cache entirely and parks it in a pool assembled from about 10% of each machine's host DRAM with a minute-level time to live. When a session resumes and that state is gone, SWA Bounded Replay reconstructs it approximately: rather than replaying layers × window tokens, it replays only the most recent window, reported at 128 tokens, and accepts that the rebuilt state is not mathematically identical to a full replay. DeepSeek says it observed little quality impact from the approximation and simulated the same behaviour during post-training. Put that together with the narrower global cache and the persistent KV footprint falls to roughly one-eighth of DeepSeek V4 Flash under identical workloads — a bigger cut than the one-quarter figure that leads the spec sheet, because it is two changes stacked rather than one format change.

Two limits sit on all of it. The report offers no end-to-end latency or cost comparison, so an eight-times-smaller persistent cache is not a demonstrated eight-times-cheaper workload: compression relieves memory pressure, and only the memory-bound part of a bill responds to that. And the saving is not something that happens to you. An engine has to implement the encoder-decoder projection, the three CSA2 modes, the FP4 cache layout and Bounded Replay before the number means anything, which is why 890 bytes is best read as a first-party specification for DeepSeek's own served API rather than a promise about a local build — the quantized conversions for llama.cpp, Ollama and LM Studio in the repository do not carry this cache with them. The headline also covers text tokens; vision embeddings sit outside it. On anything self-hosted, the number worth trusting is the one you measure. How far an engine can get with it is no longer hypothetical, though it is still not independent. vLLM implemented the decoder-side trim and Bounded Replay over the three weeks after release and published the before-and-after in the speed section below — the closest thing yet to a check on the specification, and one made by an engine measuring its own work rather than by an evaluator measuring the model.

독립 점수를 두 번 산출: 9월 가중치에서는 39, 1017에서는 45

Artificial Analysis has published the first outside evaluation of DeepSeek V4.1 Flash, and it is the number the beta could not produce. On the current version of the Artificial Analysis Intelligence Index (v4.3.2), the model scores 39 in Reasoning mode at maximum effort. On the same scale and the same setting, DeepSeek V4 Pro 0813 scores 36 and DeepSeek V4 Flash 0731 scores 34. Read together, those three rows say the thing DeepSeek's own marketing claimed and its benchmarks could not establish on their own: the smallest model in the family is now the strongest one. Artificial Analysis describes 39 as well above the median of 19 for open-weight models of comparable size. It was, for five weeks, a score for the September checkpoint alone: the evaluator's listing carried the September release date and nothing else. On October 18 that changed, and the change is the news this piece did not have when it first ran. A second row appeared for DeepSeek V4.1 Flash 1017, dated October 17, scored 45 on the same index at the same maximum effort. That puts the 1017 weights level with GLM-5.3 on 45, two points below GPT-5.6 Sol on 47 and six below Claude Opus 5 on 51 — and it makes the checkpoint the highest-scoring DeepSeek model the site has measured, five weeks after the September weights took that title. It also gives the model a standing it did not have before, and one the evaluator states outright in its October 19 write-up: second on the open-weights table, tied with GLM-5.3 and a point behind MiMo-V2.6-Pro at 46, with Step 5 Preview and Kimi K3 a point further back at 44. The six points are not spread evenly across the ten evaluations. The same write-up reports the largest moves in the agentic and knowledge components — a GDPval-AA v2.1 Elo of 1,746 against 1,600, an AA-Briefcase v1.1 Elo of 1,570 against 1,420, and an AA-Omniscience non-hallucination rate of 33% against 4%, with accuracy rising from 43% to 46% and that evaluation's index moving from -5 to +5. Those are the evaluator's own numbers for its own tests, published by it rather than reproduced here, and they describe a model that became markedly less willing to invent an answer — a different improvement from the speed one DeepSeek's changelog leads with.

The comparison that matters here is the internal one, because it is the same comparison DeepSeek's routing table makes. V4.1 Flash carries 552B total parameters with 16B active per token; DeepSeek V4 Pro 0813 carries about 1.6T total with roughly 49B active. A model with roughly a third of the backbone scores three points higher on a composite its maker did not run — nine points higher once the 1017 weights are counted, since the 1.6T flagship has not moved from 36. The size-to-score trade is the entire thesis of the Causal Encoder-Decoder design, and an evaluator with no stake in the result has now produced a number consistent with it.

The price half of the claim holds up too, and this is the first price comparison in the story that does not come from a DeepSeek rate card. Artificial Analysis lists DeepSeek V4.1 Flash at $0.30 per 1M input tokens and $1.20 per 1M output, against $1.32 and $3.96 for DeepSeek V4 Pro 0813 — 4.4× cheaper on input and 3.3× on output, with a blended $0.18 per 1M on the site's 7:2:1 cache-hit, input and output mix. It is also cheaper than the older, smaller DeepSeek V4 Flash 0731 it retires, which the site lists at $0.44 and $1.32 while scoring five points lower. Cheaper than the model it replaces, cheaper than the model it displaces, and ahead of both on the index.

Two caveats belong on the number. It is a composite of ten evaluations run at maximum reasoning effort, so it says nothing about the lower effort settings most production traffic actually uses, and nothing about how quality holds at the far end of a 1M-token window. And it is a single evaluator's score: independent, but not replicated — and the composite itself has been revised since this piece first cited it, from v4.3 to v4.3.2, which moved DeepSeek V4.1 Flash's score from 40 to 39 and its predecessor's from 35 to 34. What it does establish is narrow and useful — the one claim the beta could not support, that a 552B model is genuinely competitive with DeepSeek's own 1.6T flagship, now has a source outside DeepSeek. The 1017 row carries the same two caveats and one more. It is a single evaluator's score, for weights five weeks newer than most of the evidence here, and it is one composite rather than a replication of any vendor row. Its per-task price has also moved on the site: the September checkpoint is listed today at $0.49 per Intelligence Index task, where this piece first carried $0.27, and the 1017 checkpoint at $0.26. Treat the dollar figure as a live number on somebody else's dashboard rather than a constant — which is the same lesson the index score itself taught when it went from 40 to 39. The cost table beside the score is the rest of the caveat, and it cuts both ways. The model one place above it is also half its price: MiMo-V2.6-Pro reaches 46 at $0.13 per Intelligence Index task against $0.26 here, and Artificial Analysis's own comparison puts GPT-6.1 Sol at medium effort on 48 for $0.21. What the 1017 checkpoint does win is the head-to-head DeepSeek has been selling since September — it beats GLM-5.3-Flash on both axes of that table, 45 against 42 and $0.26 against $0.28 a task — which is the pairing where capability and cost move the same way.

다음 외부 점수: ARC Prize가 테스트를 만들고 있습니다

9월 25일, ARC Prize의 모델 테스트 책임자인 Matt Mazur는 재단 자체의 ARC-AGI-3 벤치마킹 저장소에 일급 DeepSeek 프로바이더 어댑터를 추가하는 풀 리퀘스트를 열었다. 여기에는 deepseek-v4-1-flash-low-provider-adapter라는 이름의 프로필 하나가 포함되며, 그 model 필드는 deepseek-ai/DeepSeek-V4.1-Flash이다. ARC Prize가 이 모델을 평가할 계획이라는 주장은 단 하나의 소셜 게시물로 전해졌기에, 우리는 그 게시물이 아니라 재단 자체 코드를 기준으로 확인했다. 어댑터와 프로필, 그 단위 테스트 및 문서는 diff에 들어 있으며, 그 저장소는 ARC Prize의 테스트 정책이 자사의 대화형 에이전트형 벤치마크를 위해 명시한 하네스이다. 존재하지 않는 것은 결과다. 풀 리퀘스트는 열려 있고 병합되지 않았으며, 점수도, 발표 날짜도, 어떤 추론 수준이 공개될지에 대한 진술도 없다. 중요한 것은 하나가 다가오고 있다는 소식이 아니라 측정의 형태다. ARC-AGI는 다른 리더보드들이 던지지 않는 질문을 던진다. 모델은 하네스 안에서 빠르고 저렴하고 강력할 수 있으면서도, 본 적 없는 과제에서는 60점대를 받을 수 있다. 그것이 아직 아무도 DeepSeek V4.1 Flash에 던지지 않은 질문이다.

이 프로필은 그러한 실행이 어떻게 작동할지 규정하며, 세 가지 세부 사항이 최종적으로 나올 수치를 좌우할 것이다. 이 프로필은 모델을 최대가 아니라 낮은 reasoning-effort 설정으로 실행하고, 응답을 스트리밍하며, 175,000 토큰에서 요약 압축 트리거가 있는 1M 토큰 하네스 컨텍스트를 설정하고, 출력을 262,144 토큰으로 제한한다. DeepSeek 자체 엔드포인트가 아니라 Baseten의 OpenAI 호환 엔드포인트를 통해 라우팅하며, 100만 입력 토큰당 $0.30, 100만 출력 토큰당 $1.20로 구성된다 — 이 글 뒤에서 인용된 비피크 수치가 아니라 DeepSeek의 피크 카드 요율이다. ARC Prize는 프로필에 구성된 가격을 기준으로 작업당 비용을 계산하므로, 최종 수치는 피크 요율 기준 수치가 될 것이며, 이는 테스트 당시 DeepSeek V4 Flash 0731이 기록한 작업당 $0.042와 비교하기 전에 기억해 둘 만하다. 또 다른 특이한 특징은 DeepSeek의 사고 모드 작동 방식에 대한 우회책이다: 과거 추론 내용은 요청에 도구가 포함된 경우에만 재생되므로, 어댑터는 모델이 모든 ARC 액션을 submit_action 도구 호출로 내보내게 하고, 산문에서 액션을 파싱하는 대신 추론이 첨부된 정확한 어시스턴트 메시지를 재생한다.

이것이 비교될 기준선은 ARC Prize 자체 리더보드에서 나온다. V4.1 Flash가 퇴역시키는 체크포인트인 DeepSeek V4 Flash 0731은, 그것이 속한 Flash 티어에 대해 리더보드에서 검증된 등재 항목이다: 최대 추론 노력 기준 ARC-AGI-2에서 61.4%, ARC-AGI-1에서 89.0%를 기록했으며, 태스크당 비용은 $0.042와 $0.021이다. Dots3-Note Preview는 ARC-AGI-2에서 76.8%에 위치하는데, 이는 어댑터를 지적한 게시물이 가리키는 77%이며, 같은 계정의 기대 — 훨씬 더 크고 네이티브 멀티모달인 후속 모델이 80% 미만이면 실망스러울 것이라는 — 는 공표된 목표라기보다 그의 프레이밍이다. 그 프레이밍의 나머지 부분에는 라벨이 필요하다: V4.1 Flash가 자신이 대체하는 모델보다 토큰을 29% 더 생성한다는 주장은 출처가 없고 정의되지 않았으며, DeepSeek나 ARC Prize가 공개한 어떤 것도 그것이 무엇을 측정하는지 밝히지 않는다.

One caution about which board this is. The adapter landed in the ARC-AGI-3 repository, the harness for the foundation's interactive agentic benchmark, and the separate repository that runs ARC-AGI-1 and ARC-AGI-2 carries no V4.1 entry at all — its model list has not been updated since July 27 and still names older DeepSeek endpoints. So the datable fact is that ARC Prize has wired this model into its agentic harness; whether it will also appear on the ARC-AGI-2 board, where the 61.4% comparison lives, is not something the public code shows today. The practical reading for anyone choosing a model has widened. Four independent instruments on DeepSeek V4.1 Flash are now in play — a composite index that has been run twice, once per checkpoint; a community kernel board; a harness-controlled cyber benchmark; and, since October 14, a hardware serving benchmark — and two of them put the model inside a harness on tasks it has not seen. The newer of the two is the composite, which has carried Terminal-Bench 4.0 since index v4.3: 66 terminal tasks, pass@1 averaged over three repeats, every model under the same agent harness. Its October 18 result for the 1017 weights is 33.8. The cyber benchmark, below, is the closest any of them comes, and on it this model loses the one head-to-head it makes. The ARC profile is the evaluation built to answer the reasoning question, and it is still the one without a number. Until it runs, the model's reasoning ceiling is unmeasured, and the number to hold it against is its predecessor's.

첫 외부 사이버 점수: 모델보다 하네스가 숫자를 더 많이 움직인다

9월 28일, 평가 연구소 Generality Labs는 이전 사이버 연구의 후속편으로 “사이버 벤치마크를 실행할 때 하네스가 중요한가?”라는 제목의 글을 발표했다 — 이는 DeepSeek V4.1 Flash를 동일한 사이버 벤치마크에서 다른 오픈웨이트 모델과 나란히 놓은 첫 번째로 완료된 외부 측정이다. 그 벤치마크는 Exploit Bench로, Seunghyun Lee와 David Brumley 교수가 수행한 카네기 멜런 평가이며, V8 보안 샌드박스가 활성화된 프로덕션 V8을 대상으로 에이전트가 열여섯 가지 역량의 사다리 — 취약한 코드에 도달하는 것부터, 버그를 트리거하는 것, 익스플로잇 프리미티브를 구축하는 것, 임의 코드 실행에 이르기까지 — 를 얼마나 멀리 오르는지 점수화한다. 이 연구소는 DeepSeek V4.1 Flash와 Z.ai의 GLM-5.3-Flash 두 모델을, 벤치마크 원래의 41개 과제에서 문항반응이론(Item Response Theory)으로 추출한 10개 과제에 대해 과제당 2억 5천만 토큰 예산으로 실행했고, 각 모델을 일곱 가지 하네스 구성에서 반복했다: 여섯 개의 에이전트 CLI — Claude Code, Codex, Kimi CLI, ReAct, Gemini CLI, OpenCode — 그리고 벤치마크 자체의 기본 구성으로, 글에서는 이를 Original이라고 표기한다. Original은 서드파티 CLI가 전혀 아니다: 토큰 제한이 없고, 300턴 상한이 있으며, 컨텍스트 압축이 없다. 이 결과를 제기한 계정은 Original을 대략 파이썬 while 루프라고 묘사한다; 그것은 연구진의 설명이 아니라 그 사람 자신의 특징 규정이며, 연구진 자신의 설명은 위에 적힌 바로 그 설명이다.

이 결과는 단일 점수로는 드러나지 않았을 격차다. 평균 역량은 16점 만점이다. DeepSeek V4.1 Flash는 Claude Code에서 9.3으로 최고이며, Codex가 8.9로 2위다. 같은 모델은 Kimi CLI에서 7.4, OpenCode에서 5.9, Gemini CLI에서 5.7, ReAct에서 5.3, Original에서 4.7로 떨어진다. Original은 자체 벤치마크 기본값이며, 일곱 가지 중 최악이다. GLM-5.3-Flash는 양쪽 끝에서 그 순위를 뒤집는다. 최고 구성은 Codex에서 10.6으로, 연구 자체 데이터셋이 전체 41개 과제 세트에서 Mythos Preview에 대해 제시하는 10.02보다 높다. 다만 연구진은 10개 샘플 하위 집합이 완전히 대표적이지는 않다고 경고한다. 최악은 Claude Code에서 6.2인데, 이는 DeepSeek V4.1 Flash의 최고 구성과 같은 하네스이며, Gemini CLI에서는 8.9, Kimi CLI에서는 8.2, OpenCode에서는 8.1, ReAct에서는 7.7, Original에서는 6.4이다. 일곱 가지 전체 평균으로 보면, GLM-5.3-Flash는 약 8.0, DeepSeek V4.1 Flash는 약 6.7에 위치한다. 그래서 이 결과와 함께 퍼진 주장의 두 부분 모두 공개된 데이터와 맞아떨어진다. 하네스 선택은 몇 점의 가치가 있고, 이 벤치마크에서 DeepSeek V4.1 Flash는 두 모델 중 더 약한 쪽이다. Claude Code에서는 3점 넘게 앞선다는 점만 빼면.

연구자들은 그 격차를 완전히 설명할 수 없다고 분명히 말한다. 그들이 가진 하나의 유력한 단서는 각 하네스가 컨텍스트를 압축하는 방식이며, 그 차이는 문구의 문제다. ReAct의 압축 프롬프트는 모델에게 “당신(또는 당신 자신의 또 다른 인스턴스)이 작업을 재개할 수 있게 해 주는 후속 요약을 작성하라”라고 요청한다. 연구소가 공개한 기록에서 이는 때때로 모델이 작업이 끝났다고 선언하게 만들고, 그 뒤 하네스가 다시 모델을 깨우면 예산이 바닥날 때까지 동일한 보고서를 수십 번 다시 제출한다. Codex는 같은 순간을 “작업을 이어받을 다른 LLM을 위한 인계 요약을 작성하라”라고 규정하며, 연구소는 다른 에이전트에게 말을 거는 것이 모델이 긴 예산에 걸쳐 계속 진전을 이루게 하는 요인이라고 가설을 세운다. 그들은 이것을 원인이 아니라 가설이라고 부른다. 표를 해석할 때 중요한 구성 오류도 하나 있다: ReAct의 압축은 의도된 트리거 대신 100K 토큰으로 설정되어, 설계가 요구한 것보다 훨씬 자주 압축했고, 이로 인해 두 모델의 ReAct 행이 낮아진다. 이 연구 자체의 주의사항은 여기서 나온 어떤 수치와도 함께 기억할 가치가 있다 — 모델 두 개, 하네스 구성당 샘플 10개, 벤치마크의 원래 41개 과제에 대한 모델당 질문 샘플 70개, 그리고 오류가 발생했거나 끝나지 않은 샘플 세 개(그중 두 개는 GLM-5.3-Flash)다. 저자들은 작은 표본은 본질적으로 잡음이 많고, 그 결과가 다른 영역과 과제로 확장된다고 가정해서는 안 되며, METR의 선행 연구는 장기 지평 과제에서 하네스 간에 통계적으로 유의미한 차이를 발견하지 못했다고 밝힌다. 이것은 벤치마크 소유자가 아니라 다른 누군가의 벤치마크를 실행한 외부 연구소의 결과이며, 동료 검토를 거치지 않았다. 이것이 벤더 수치가 아니라는 점은 분명하며, 이 글에서 벤더 자체의 설명을 단지 뒷받침하지 못하는 데 그치지 않고 그것과 모순되는 첫 번째 외부 수치다.

여기에는 라우터가 작용할 수 있는 한 가지와 그렇지 않은 한 가지가 있습니다. 라우터는 여러분의 하네스를 선택할 수 없습니다. 하네스란 여러분이 이미 실행 중인 에이전트 루프이며, 이 연구의 핵심은 그 선택이 결정적이라는 점입니다. 라우터가 할 수 있는 일은 그 쌍에서 모델 쪽을 한 줄 변경으로 만드는 것입니다. DeepSeek V4.1 Flash와 GLM-5.3-Flash는 둘 다 단일 OpenAI 호환 키로 OrcaRouter를 통해 라우팅할 수 있으며, 제공자 자체의 정가에 0% 마크업이 더해진 가격입니다. 따라서 자체 모델-하네스 조합을 찾고자 하는 팀은 두 번째 계약이나 두 번째 통합 없이 고정된 하네스 아래에서 모델을 교체할 수 있고, 교체가 회귀인 것으로 드러나면 그 뒤에 검증된 모델을 자동 페일오버로 남겨 둘 수 있습니다. 이는 아래 도입 섹션에서 제시하는 것과 같은 논지를, 하네스 연구가 방금 구체화한 결정에 적용한 것입니다. 이 증거에 따르면 여러분이 실행하는 하네스는 여러분이 고르는 모델보다 더 중요할 수 있으며, 그 쌍에서 모델 쪽은 여러분이 오후 한나절 만에 바꿀 수 있는 쪽입니다.

프런티어와 비교해 어디에 위치하는지, 그리고 여전히 뒤지는 지점

The September model card compares V4.1 Flash against Opus-5.0, GPT-5.6 Sol, Kimi K3 and GLM-5.3 at maximum reasoning effort, with its own DeepSeek V4 Pro and DeepSeek V4 Flash as internal baselines. Every figure below is vendor-reported, run on DeepSeek's harnesses, and unreproduced by any third party. The card notes that agentic results use the Minimal mode of DeepSeek Harness with a 1M-token context for code benchmarks and a 512K window for visual ones, so these are not directly comparable to published third-party numbers on the same benchmarks. The independent scores in the sections above are different measurements on different sets of evaluations; they support the general picture and corroborate no individual row below, and the October 17 rows in the section above come from a different card and a different run of the same benchmarks.

에이전트 및 코딩 작업에서, 벤더가 제시한 수치는 V4.1 Flash를 자사가 이름을 밝힌 프론티어 모델들보다 위에 둡니다:

• Terminal-Bench 2.1 — 90.6, Opus-5.0의 89.1, GPT-5.6 Sol의 88.8, DeepSeek V4 Pro의 87.9에 비해

• DeepSWE v1.1 — 74.2 해결, 74.0, 73.0, 62.7 대비

• CyberGym — 88.1로, GPT-5.6 Sol과 GLM-5.3의 84.5, 그리고 DeepSeek V4 Pro의 83.3과 대비된다

• AutomationBench — 54.8, Opus-5.0의 50.3과 DeepSeek V4 Flash의 37.7 대비

• Agent's Last Exam — 28.6 및 25.7 대비 31.8

• Codeforces — 3,471 레이팅으로, 3,348의 DeepSeek V4 Pro와 3,289의 DeepSeek V4 Flash보다 앞서 있습니다

• MathArena Apex — 65.6, Kimi K3와 동일한 수준이며 65.3을 기록한 DeepSeek V4 Pro를 앞선다

그리고 더 어려운 에이전틱 스위트에서는 그렇지 않습니다:

• Terminal-Bench 3.0 — 30.0, Opus-5.0의 43.3과 GPT-5.6 Sol의 34.4에는 훨씬 뒤처지지만, DeepSeek V4 Pro의 11.8보다는 훨씬 앞선다.

• Terminal-Bench 4.0 — 31.2, 51.8과 39.9에 맞서, 그리고 12.4에서 다시 DeepSeek V4 Pro를 앞서

• ProgramBench — 20.3, Opus-5.0의 37.0과 대비

• NL2Repo-Bench — 64.0, Opus-5.0의 75.3 대비

• HLE — 36.8, 56.3 및 44.5와 비교

• ExploitGym — 15.3, 33.7 및 22.1 대비

The shape of that list is more informative than any single row. V4.1 Flash clears the frontier on benchmarks that reward sustained tool use inside a harness, and loses on the ones that reward depth of reasoning — the pattern you would expect from a model engineered around throughput and context economics rather than around raw reasoning ceiling. The internal comparison is the striking one: it beats DeepSeek V4 Pro on essentially every row, including the rows where both lose to Opus-5.0. DeepSeek is not claiming its cheapest model is the best model on the market. It is claiming its cheapest model is the best model DeepSeek makes, which is a claim its own routing table backs up.

비용은 얼마인지, 그리고 Pro 비교는 실제로 어떤 모습인지

September 10's rate card is now simply the V4.1 Flash card, published on DeepSeek's Models & Pricing page with a peak and off-peak split:

• 입력, 캐시 적중 — 100만 토큰당 비수기 $0.003, 피크 시간 $0.006

• 입력, 캐시 미스 — 100만 토큰당 비수기 $0.15, 피크 시 $0.30

• 출력 — 100만 토큰당 오프피크 $0.60, 피크 시 $1.20

• 피크 시간대 — UTC 기준 월요일~금요일 01:00–04:00 및 06:00–10:00; 그 외 모든 시간은 피크 요금의 절반으로 청구됩니다.

• 위안화 기준 — 비수기에는 ¥0.02, ¥1, ¥4이며, 성수기에는 두 배인 ¥0.04, ¥2, ¥8

Compare that with the model it is displacing. DeepSeek V4 Pro bills $0.022 per 1M cache-hit input, $0.66 cache-miss input and $1.98 output off-peak — roughly 4.4× on cache-miss input and 3.3× on output. Those are the numbers the September 14 rerouting would have applied to every deepseek-v4-pro call, which is why the plan was a price story as much as a capability story, and they are still the reason the comparison matters: DeepSeek withdrew the reroute, not the rate card. Artificial Analysis reaches the same arithmetic from its own listings — $0.30 and $1.20 per 1M for V4.1 Flash against $1.32 and $3.96 for DeepSeek V4 Pro 0813, the same 4.4× and 3.3×, and a blended $0.18 per 1M on the site's 7:2:1 cache-hit, input and output mix. The card also undercuts DeepSeek V4 Flash 0731, the model it retires, which the same site lists at $0.44 and $1.32.

토큰당 가격은 작업당 비용이 아니며, 베타가 바로 그 증거입니다. 첫날 테스터들은 5분 세션에서 새 모델이 DeepSeek V4 Flash에서 같은 실제 시간을 보냈을 때보다 잔액을 더 빠르게 소진했다고 보고했습니다. 이 모델은 토큰을 빠르게 출력하도록 만들어졌고 청구 금액은 분이 아니라 토큰에 비례하기 때문입니다. Pro 가격의 4분의 1인 카드는 작업이 비슷한 토큰 수로 끝난다면 이득이고, 모델이 거기에 도달하기 위해 더 오래 생각한다면 본전입니다.

Where an outside number does exist on that question, it points both ways. Artificial Analysis puts DeepSeek V4.1 Flash at $0.27 per Intelligence Index task on its own evaluation set — a real per-task figure, produced outside DeepSeek, and a low one. But the same run generated 250M output tokens against a median of 140M for comparable models, and the site's summary calls the model very verbose in the same breath as calling it fast. Verbosity is free inside a benchmark harness and expensive in production, which is precisely the mechanism the beta testers were describing when they reported balances draining faster in wall-clock terms. The honest state of the per-task question is that one independent evaluator has now measured both halves of it, on its own evaluation set, and the answer still depends on what you are running. The composite cannot show where the extra output goes, and the beta testers never said either. The first hands-on reports do, and they are in the speed section below: the model is verbose because it is checking. The October 17 update is the first time DeepSeek has claimed to move that number: the changelog says the new checkpoint “reduces task runtime and output length,” and the post a few hours later says the same thing in fewer words — “faster task completion, and shorter outputs.” That is DeepSeek addressing, on its own authority, the exact quantity the outside measurement complained about, with no figure attached to either statement. The first outside look at it, on October 18, found the aggregate volume down — 210 million output tokens across the index against 250 million for September — and the per-task figure flat, at 89,495 tokens against 88,574.

한 달 비용: 80달러 허용량, 그다음 무료 등급

위의 카드는 토큰 비용이 얼마인지에 답한다. 다른 수치가 두 번 움직였는데, 두 번 모두 OpenCode에서 나왔다 — 이 코딩 에이전트의 Go 구독은 이 모델에 대한 접근을 종량제 토큰이 아니라 월간 사용 허용량으로 가격을 매긴다 — 그리고 그것은 모델이 아니라 자사 구독에 대해 말한 것이었다. 9월 25일, 이 플랫폼의 계정은 이렇게 게시했다: "Operation Cheepseek Phase 2: DeepSeek v4.1 Flash에 대한 $60 상당의 사용량이 이제 영구화되었습니다." 10월 12일에는 같은 형식으로 속편을 게시했다 — "Operation Cheepseek Phase 3: DeepSeek V4.1 Flash에 대한 $80 상당의 사용량" — 그리고 이 모델에 대해 9월 말에는 $60이었던 자사의 가격 페이지에는 $80이 적혀 있었다. 그 확인은 10월 13일에 같은 페이지를 대상으로 두 번째로 이루어졌고, 동일한 행을 반환했다: 월 $10인 Go 등급에서 $80, Go Plus에서 $160. 따라서 이 수치는 한 게시물이 스스로에 대해 주장한 것이 아니라, 한 플랫폼이 이 모델을 취급하는 데 대해 공개한 가격이며, 하루 뒤에도 여전히 그 자리에 있었다. 두 게시물을 함께 읽으면 "영구"가 결국 무엇을 의미했는지 말해준다: 만료일이 없다는 것 — 그리고 3주 뒤에는 더 높아진 숫자라는 것. 그 변화가 의미하는 것은 모델에 접근하는 새로운 방식이 아니다 — 그것은 3일 뒤, 아래 문단에서 등장한다. DeepSeek V4.1 Flash는 9월 10일부터 DeepSeek 자체 API에서 응답해 왔고, OrcaRouter는 출시 이후 DeepSeek 자체 요율로 이를 취급해 왔으므로, 이 글에 나오는 가중치와 아키텍처, 그리고 모든 벤치마크 행은 구독 가격 재조정의 영향을 받지 않았다. 10월이 바꾼 것은 누가 이 모델을 취급하려 하는지, 그리고 어떤 조건으로 취급하는지이다.

무료 항목은 앞의 두 가지와는 다른 종류의 변화이며, 이 셋 중에서 이번 주에 독자가 할 수 있는 일을 바꾸는 유일한 변화다. 10월 15일 OpenCode는 이 모델 — DeepSeek V4.1 Flash Free, 모델 ID deepseek-v4.1-flash-free — 을 위한 별도 항목을 추가했고, 이는 무료로 제공된다. 이 플랫폼의 공개 모델 목록은 이를 종량제 ID 옆에 함께 표시하고, Console 문서에는 이를 한시적으로 무료라고 설명하며 제공자가 데이터를 학습에 사용하지 않는 제로 리텐션 정책을 따른다고 밝히는 무료 모델 항목이 있으며, 공개된 가격표는 입력, 출력, 캐시 읽기를 모두 Free로 표기한다. 다만 이것은 완전한 제품이 아니다. 무료 항목의 카탈로그상 제한 — OpenCode 자체 모델 목록의 출처인 카탈로그 models.dev에 10월 15일 추가됨 — 은 종량제 항목을 기본 모델로 지정하고 컨텍스트 창을 200,000 토큰, 출력 상한을 32,000 토큰으로 제한하는데, 이는 종량제 deepseek-v4.1-flash 항목이 제공하는 1M 컨텍스트와 384K 최대 출력과 대비된다. 이는 더 작거나 증류된 모델이 아니라 동일한 가중치에 더 좁은 제공 조건을 붙인 것이며, 출력 상한은 에이전트 작업에서 가장 먼저 걸리는 조항인데, 파일 하나를 편집하는 것만으로도 이를 넘길 수 있기 때문이다. 플랫폼은 이 arrangement에 종료일을 붙이지 않았고 이에 대한 요금표도 공개하지 않았는데, 요금 자체가 없기 때문이다. 이 모든 것은 DeepSeek에서 나온 것이 아니다. 이는 자사가 소유하지 않은 모델에 대한 한 플랫폼의 상업적 결정으로, 그 플랫폼 자체의 모델 목록, 자체 문서, 자체 가격표에 명시되어 있으며 이 모든 정보의 유일한 출처다. 우리는 무료 엔드포인트가 부하 상황에서 어떻게 작동하는지에 대한 독립적인 측정을 찾지 못했고, “한시적”은 일정이 아니라 플랫폼의 표현이다. 이것을 오늘 이 모델 앞에 200,000 토큰 창을 두는 가장 저렴한 방법으로 읽어야 하며, 모델 자체가 무엇인지에 대한 변화로 읽어서는 안 된다.

The mechanics come from that pricing page, read on October 12 and re-read on October 13. Go is a $10-per-month subscription, and its table pairs every model on it with an estimated request count per five hours and a monthly usage limit, in separate columns for the two tiers. DeepSeek V4.1 Flash carries the largest allowance on the page: $80 a month and 34,670 estimated requests per five hours on Go, doubling to $160 and 69,330 on Go Plus. The rows bill the model at DeepSeek's own card — $0.15 per 1M input tokens, $0.60 output and $0.003 for a cached read at off-peak, doubling to $0.30, $1.20 and $0.006 inside the weekday peak windows — and those are the vendor's figures, unchanged since the September reading this section already carried. Two things the page showed then are not on it today, and both are worth flagging because the earlier reading relied on them: the three-bucket draw-down rule — 20% over any five hours, 50% across a week and 100% across a month — no longer appears, and the documentation URL the request estimates were originally read from, opencode.ai/docs/go, returned 404 when we checked. The request counts above are the platform's arithmetic on a call shape it does not publish, not ours.

The change is checkable against the page, and so is the fact that there has now been more than one. In late September the same rows read $60 with no expiry attached, after a promotional $15 allowance was multiplied four times to $60 and carried September 20 as the date the multiplier would lapse; the request estimates were published in two columns, 6,500 / 16,250 / 32,500 standard against 26,000 / 65,000 / 130,000 with the multiplier applied. The rows read $80 on both readings, and the published estimates for this model are 34,670 per five hours and 69,330 a month — a higher dollar allowance sitting beside a lower monthly request count than September's expanded set, which cannot both describe the same assumed call shape. The page states no token mix and offers no explanation, so the difference is a fact about the two readings rather than something we can resolve from outside. The rates themselves were never a discount on DeepSeek's card — per-token pricing has been the vendor's throughout, and it still is. What an allowance does not do is say anything about quality: it is one platform's judgment that it can carry this model's traffic at those rates without a deadline, which is evidence about the serving ecosystem rather than about the model.

The distinction matters when you are choosing, because a flat ceiling and a pass-through meter are different products. DeepSeek V4.1 Flash is on OrcaRouter at DeepSeek's own rate with 0% markup — $0.15 per 1M input and $0.60 per 1M output off-peak, the same numbers the card above lists, with nothing bundled around them: no subscription to buy and no monthly allowance to exhaust. An $80 ceiling on a fixed monthly price is the right shape for one developer running an agent all day, where the question is what the month costs — and a ceiling already raised once, by a third, is also a reminder that the figure is a commercial decision rather than a property of the model. A free tier with a 200,000-token window is the right shape for trying the model this week and the wrong shape for the codebases that need the full window, which is the one thing the two tiers of the same model do not share. A vendor rate passed through unchanged is the right shape for a service, where the question is what each call costs and whether the provider's next price move reaches you the day it happens. Both answer what this model costs; they answer it for different spenders.

속도: 베타의 수치, 제작자의 주장, 그리고 엔진 자체의 전후 비교

베타가 문을 열었을 때, 처리량이 핵심 화제였습니다. 개발자들은 긴 생성에서 초당 약 420개의 출력 토큰을 보고했고, 피크는 500을 넘기도 했으며, 널리 공유된 한 실행에서는 3분 이내에 71,000개 이상의 토큰을 생성했습니다. 이들은 여전히 통제되지 않은 조건에서 이뤄진 커뮤니티 측정치이며, 여전히 그렇게 읽어야 합니다. 하지만 그 배후의 메커니즘은 이제 추측이 아니라 문서화되어 있습니다. 토큰당 890바이트의 KV 캐시와 16B 파라미터를 활성화하는 디코드 경로는 처리량을 위해 만들어진 설계이므로, 이 아키텍처는 적어도 그런 수치를 내기에 알맞은 형태입니다. 그 수치가 실제로 얼마나 큰지는 다음 문단에서 다룰 문제이며, 답은 더 작습니다.

DeepSeek's own claim, in the changelog, is that V4.1 Flash “has comprehensively surpassed V4 Pro in performance, cost, speed, and total time.” The word doing the work there is total time: it is the vendor framing speed as task completion rather than tokens per second, which is the metric that survives contact with a real agent loop that writes a file, reads the error and tries again. Artificial Analysis now puts a measured figure on the older claim: 185 output tokens per second for DeepSeek V4.1 Flash, against 175 for DeepSeek V4 Flash 0731. That is a real third-party number and it is roughly half the 420 tokens per second developers reported during the beta, which is why the beta figures should stay where they belong — in the uncontrolled-conditions column.

On October 7 the vLLM project published its accounting of the three weeks since release, and it is the first measured speed figure in this piece that is neither a token counter nor a vendor slogan. Writing as Inferact and the vLLM Team, the project reports that DeepSeek V4.1 Flash on its current build runs 1.9× faster than its own day-0 implementation at low concurrency and delivers 5.3× the throughput under a 150 TPS constraint — the figure its summary carries, against “about 5×” in the body of the same post. Both are measured on the SemiAnalysis AgentX agentic serving benchmark, with TP4 and FlashInfer attention behind the low-latency number and DEP2 with data-parallel attention behind the throughput one, on NVIDIA GB200 and GB300 hardware. Three things about that result are worth keeping straight. It compares vLLM against vLLM: the baseline is the project's own day-0 recipe, not another engine and not another model, so the number says how much the serving stack improved rather than how this model compares with anything else. It is the runtime's own account of its own work, and no third party has reproduced it. And the largest single gain comes from the runtime rather than the weights — SWA bounded replay, which reruns only the last 128 tokens of a sliding window instead of replaying that window across layers, now ships on by default for this model behind a single flag. The project reports roughly 30% off time to first token from replay plus CUDA-graph capture, a 30–40% cut in prefill computation from the decoder-side trim, and nearly 70% off TTFT at around 100K throughput once three optimisations stack. It also reports where replay hurts: without CUDA graphs it is slower than the baseline on short prompts, up to 12% worse at 1K tokens, because launch overhead outweighs the work saved. Accuracy was checked on GSM8K and GPQA with no meaningful difference reported, gaps within about 1.5 standard errors — the figure to hold against the report's own statement that bounded replay is approximate rather than exact. The project adds that the model needed no KV-cache offloading anywhere in the benchmark and expects offloading to start paying only at higher concurrency with prefill and decode split across machines. This is a different measurement from the kernel board below, which scores kernels the model writes; vLLM's figures are about what the model costs to serve once an engine has implemented its architecture, and the engine here is the one reporting. The project posted again on October 17, the day DeepSeek swapped the checkpoint, and the number it led with was a different kind of figure: 1.3B tokens per dollar of total cost of ownership on SemiAnalysis AgentX, at 830,000 tokens per second per GPU counting cached, input and output tokens together, while holding 100 output tokens per second for every user. That is a cost-of-serving result rather than a speed delta, and the serving-cost section below reads it in full. Three things about it bear on this paragraph: it is the same project, the same model and the same benchmark three weeks on from the October 7 accounting; the project frames that accounting as chapter one, at 5x the agentic throughput of day zero, with a next chapter still to come, and no second DeepSeek post has appeared on its blog since; and a cost curve is still the runtime's side of the ledger, not an outside verdict on the model. That distinction is also why this number matters to someone who will never run a node. None of the work is work a caller does: a serving stack that gets 1.9× faster at low concurrency makes the requests it handles faster, and reaching this model through an API is how most teams will benefit from that without owning the hardware. DeepSeek V4.1 Flash is on OrcaRouter at DeepSeek's own rate with 0% markup, on the same OpenAI-compatible key as the rest of the line, with a proven model behind it as automatic failover — the model and the pricing move at the provider's pace, and the choice of engine stays on the provider's side of the line.

서빙에 드는 비용: 최초의 독립적 하드웨어 수치

위의 모든 내용은 모델을 측정한 것입니다. 10월 14일 Artificial Analysis는 DeepSeek V4.1 Flash를 AA-AgentPerf에 추가했습니다. 이 벤치마크는 모델이 아니라 가속기 시스템을 대상으로 실행되며, 그 결과는 요금표로는 해결할 수 없는 질문에 대한 첫 공개 답변입니다: 데이터센터 전력 1메가와트가 이 모델을 얼마나 감당할 수 있는가. AA-AgentPerf는 기록된 코딩 에이전트 세션 — 다중 턴 추론, 도구 호출, 100K 토큰을 넘어서는 컨텍스트, 프롬프트 캐시 적중 포함 — 을 라이브 서빙 스택에 재생하고, 각 에이전트가 여전히 에이전트별 속도 목표를 충족하는 상태에서 시스템이 유지하는 최대 동시 에이전트 수를 찾아냅니다. 이 모델의 경우 핵심은 NVIDIA의 랙 스케일 부품입니다. 사용자당 초당 100 출력 토큰에서 GB300 NVL72가 메가와트당 초당 약 2억 6,100만 토큰으로 선두를 달리고 있으며, B300이 약 2억 2,000만, B200이 1억 9,300만, GB200이 1억 7,800만, AMD의 MI355X가 1억 2,800만으로 뒤를 잇습니다. 이는 Artificial Analysis의 수치로, 오늘 데이터센터 추론 페이지에서 확인한 것이며, 벤더의 수치가 아니라 특정 하드웨어 및 서빙 구성에 대한 자체 측정값입니다. 두 가지 세부 사항은 기억해 둘 가치가 있습니다. 메가와트당 수치는 실행 중 측정된 가속기 전력 — GPU 다이 + HBM — 에서 계산된 것이며, 정격 TDP에서 계산된 것이 아닙니다. 그리고 차트의 각 점은 그 뒤에 있는 엔진, 버전 및 숫자 형식을 명시하므로, 그 수치는 서빙 스택에 대한 주장이며 가중치에만 관한 것이 아닙니다.

The reason this model's chart looks unlike the others is its size. DeepSeek V4.1 Flash fits on as few as two accelerators, which is why a rack-scale system gets less of an advantage here than it does on a larger model: Artificial Analysis reports B300 staying within about 20% of GB300 NVL72 across the curve and leading it at both ends — the most throughput per megawatt when per-user speed is allowed to fall, and the fastest single agent when it is not. At batch size one, every Blackwell part and the MI355X land between roughly 600 and 670 tokens per second for a single user, against 372 to 392 on H100 and H200. Set against DeepSeek V4 Pro 0813 on the same axis, the smaller model returns about 1.3× the throughput per megawatt on GB300 and about 6.2× on B200 at the 100 tok/s mark — the efficiency claim the architecture section makes from the KV-cache side, arriving from a third party that measured a system rather than a specification. The configurations behind those numbers show the design's central trick in production rather than only on paper: on both the B300 and the MI355X two-accelerator runs, the 196B Engram table is served out of pinned host memory through vLLM's CPU-offload path instead of sitting on the accelerator — the mechanism the release notes described, now visible in the setup behind a published measurement. A second cost figure, in a different currency, arrives from the benchmark the October 7 vLLM post was measured on. On October 17 the vLLM project published a chart putting DeepSeek V4.1 Flash at 1.3B tokens per dollar of total cost of ownership on SemiAnalysis AgentX, with separate curves for GB300 NVL72 under Dynamo, GB200 NVL72, B200, H100 and MI300X. The cost tier behind it is labelled owning at large-hyperscaler volume, and the chip-hour assumptions underneath run from $0.95 for an MI300X and $1.17 for an H100 to $1.73 for a B200 and $2.31 for a GB300, drawn from SemiAnalysis's own July 2026 pricing survey. Two limits belong with it. It is the vLLM project reporting SemiAnalysis's cost model, and neither party has had it replicated. And the chart is dated October 16, the day before DeepSeek replaced the weights behind the model name, so it too describes the September checkpoint. That makes three currencies in this piece for one model: tokens per second per user, tokens per second per megawatt, and tokens per dollar. The last is the one a buyer without a datacenter can act on, because it is the quantity a provider's cost improvement eventually reaches them as. It remains a serving measurement and not a quality one: nothing here says how good the model is, and one leaderboard's hardware results are not a replication of any benchmark row above. What it does establish is that the cost side of this model's rate card now has outside evidence under it. A model a provider can serve at several times the throughput per megawatt is a model a provider can afford to keep cheap, and DeepSeek V4.1 Flash is on OrcaRouter at DeepSeek's own list price with 0% markup, which is the shape that passes a provider's cost improvement through instead of holding it.

두 공급업체의 실리콘에서 이 엔진에 드는 비용은 얼마인가

Until this week the throughput question about this model was answered only by other people's measurements — a runtime accounting for its own work, a benchmark site measuring a system, developers counting tokens in a beta. DeepSeek opened DeepInfer, the framework it says powers this model's production service, under Apache-2.0 this week — the repository appeared on October 16 and the first commit landed on October 20 — and shipped with it a decode Pareto plot for the 1017 checkpoint measured on 64 SM100 accelerators inside one scale-up domain. The configuration printed on the plot is the part that matters, because it is the configuration the comparison below turns on: expert parallelism of 64 over data parallelism of 64, the “384+64 experts” the plot prints, a 131,072-token context, DSpark draft length fixed at three tokens, and a per-token acceptance rate of 75%. The curve tops out just under 30,000 output tokens per second per accelerator at its least interactive end, falls through the mid-20,000s as per-user speed rises past 100 tokens per second, and is down to a few thousand by the chart's right edge. It is DeepSeek measuring its own engine on its own model, so it is not an independent number — but it is the first decode curve for this architecture published by the lab that built it, and the sweep script that produced it is in the same repository.

The write-up published beside the code supplies the operating point, and it is the reason the numbers below are quoted at 230 output tokens per second rather than at some abstract maximum. DeepSeek's own production figures put decode at 14K tokens per second per accelerator while the service holds 230 output tokens per second per user — 300 in coding workloads — against a 96% input cache hit rate, an average input of 77K tokens, a 400 ms mean and one-second p99 time to first token, and prefill provisioned at close to a 1:1 ratio with decode, across a single cluster it says serves 155 trillion tokens a day. The revenue line beside those figures, $38 per accelerator-hour at off-peak pricing, is the first cost number DeepSeek has attached to serving this model publicly. None of it is audited, all of it is the vendor describing its own service, and it is still the most concrete picture of what this architecture costs to run that exists.

The other half of the comparison comes from Huawei's silicon, and it is the weaker-sourced half. A benchmark sheet dated September 30 and headed “Ascend 950DT DeepSeek-V4.1-Flash Benchmark” reports the model deployed across 32 chips at FP8–FP4 mixed precision, with data-parallel attention and expert-parallel mixture-of-experts, on a 128K-token sequence, using DSpark speculative decoding with a forced expert-parallel load-balancer configuration and an average accepted length of 4.1 tokens. It carries three points. Its latency column is time per output token, so the reciprocal of each entry is that row's per-user speed, and that is what lets the sheet be laid against the curve:

• 글로벌 배치 1,536, 출력 토큰당 9.41ms — 사용자당 초당 약 106 출력 토큰 — 칩당 초당 5,102 토큰.

• 글로벌 배치 384, 출력 토큰당 4.40ms — 사용자당 초당 약 227 출력 토큰 — 칩당 초당 2,727 토큰.

• 전 세계 배치 2,560, 출력 토큰당 13.79ms — 사용자당 초당 약 73 출력 토큰 — 칩당 초당 5,800 토큰으로, 이 표에서 가장 높은 수치이자 헤드라인에 사용된 값입니다.

Laid against the NVIDIA curve, the middle two rows are the ones that bite, because they sit at the speeds a service actually runs at. The thread that surfaced the sheet reads the comparison as 25,500 tokens per second per accelerator against 5,102 at 100 output tokens per second per user, and 20,000 against 2,727 at 230 — a gap of five times at the first operating point and 7.4× at the second. Reading the same published curve directly moves those NVIDIA figures around: a pixel pass over the image lands nearer 28,000 at the 100-token mark and nearer 17,000 at 230, on either side of the thread's numbers; the band they give is about five and a half times at the first operating point and about six at the second, which narrows the thread's range without changing its shape. That is what a benchmark read off a chart is worth, and it does not change the direction of the result. It is also worth noting that the sheet's 5,800 headline is its least interactive row — the point where NVIDIA's curve is closest to its own peak, and where the ratio is back to about five.

I could not find that sheet on DeepSeek's or Huawei's own pages; it reached this piece as an image attached to a thread on X, and the three Ascend figures should be read as an unpublished vendor benchmark until somebody posts the run behind them. What is checkable is the asymmetry in the configurations, and that comes from DeepSeek's own documentation. DeepInfer's deployment notes recommend prefill at expert parallelism of 2 to 8 and decode at 64 to 128, on “ordinary 8-GPU machines for prefill, ordinary 64–128-GPU supernodes for decode.” On NVIDIA that is a single NVLink domain, which is what the plot above was measured inside. On Ascend, DeepSeek's own communication library says plainly what is not finished: “Larger EP sizes and combine remain under optimization,” with combine carrying extra local reduction overhead and HBM contention against the URMA path, and Huawei firmware work still to come — a caveat public in DeepEP-Ascend since its first release on September 29. That is the fair way to read the gap: the NVIDIA number comes from the shape DeepInfer was tuned for, and the Ascend number from a stack whose own author says the wide-EP path is still being worked. Some of five to seven times is the chip, some of it is the interconnect and the software under it, and nobody has separated the two.

What the gap is worth in money depends on prices nobody has published, and the thread runs the arithmetic anyway: at roughly $139,000 per Blackwell accelerator, extrapolated from grey-market rack pricing, against a rumored $37,000 for a 950DT, NVIDIA has to deliver about 3.8× the throughput per chip to match Ascend on acquisition cost alone; the thread puts the crossover near 6,800 tokens per second per chip. The published Ascend points sit below that line; the thread's forecast that the wide-EP path lifts them to 8,000–10,000 sits above it, and that is a forecast rather than a measurement. Every input is soft — a grey-market price, a rumored one, and figures read off a curve — and none of it accounts for networking, power, cooling, maintenance, or the value of being able to buy the hardware at all.

For anyone who will never rack a node, the useful part is not the scoreboard. It is that the cheapest model in DeepSeek's line is now documented as cheap to serve on two vendors' silicon, with the vendor's own engine published and the numbers behind it downloadable — which is the kind of evidence a price rests on. It changes no rate card: DeepSeek V4.1 Flash is on OrcaRouter at the provider's own list price, $0.15 per 1M input tokens and $0.60 per 1M output at off-peak rates, passed through with 0% markup, so when a provider's serving cost falls the saving arrives as a price that does not have to move rather than as a resale margin. A serving cost argued over on two vendors' hardware is a price with more room under it than one nobody has measured at all.

커널 엔지니어링: DeepSeek의 것이 아닌 첫 번째 스코어보드

위의 벤치마크 행들은 DeepSeek 자체 하네스 수치이며, 현재 존재하는 외부 점수들은 저수준 프로그래밍이 아니라 일반 추론과 사이버 익스플로잇을 측정한다. 이 모델의 커널 엔지니어링에 대한 첫 제3자 검토도 존재하는데, 그것은 DeepSeek이 의뢰하지 않은 형태로 등장했다: 모델에게 실제 CUDA 커널을 작성하게 하고 그 결과를 하드웨어의 이론적 상한인 루프라인에 견주어 점수를 매기는, 커뮤니티가 운영하는 에이전트형 벤치마크다. 리더보드는 RTX PRO 6000에서 실행되는 kernelbench.com이며, 두 벤더 모두로부터 독립적이다. 그 감사 방법론은 동료 검토를 거친 것이 아니라 커뮤니티가 보고한 방식이므로, 그 백분율은 인증된 측정값이 아니라 방향성을 보여주는 제3자 신호로 읽어야 한다. 또한 이 이야기에서 DeepSeek의 수치가 직접 경쟁자의 수치와 나란히 놓이는 첫 번째 스코어보드이기도 하다.

헤드라인 행은 이 아키텍처가 중심에 두고 구축된 바로 그 행이다. DeepSeek V4.1 Flash 자체의 희소 어텐션 — 토큰당 890바이트 캐시를 생성하는 CSA2 스택 — 은 모델이 CUDA 커널로 작성했으며, dense 등가 루프라인의 50.19%에 도달해 리더보드에서 Claude Fable 5.1(106.27%), Claude Opus 5(103.67%), Claude Fable 5(72.66%)에 이어 4위를 차지했다. 먼저 눈에 띄는 것은 제작자의 자체 아키텍처가 그 모델이 가장 잘하는 커널이라는 점이다: 희소 어텐션 설계로 훈련된 모델은 상위 3개 아래의 리더보드에서 다른 어떤 항목보다 희소 어텐션 커널을 더 효율적으로 작성한다. 두 번째는 제보자가 지적한 비교다. 같은 커널에 대한 GLM-5.3 Flash의 시도는 루프라인의 4.45%에 도달한다 — 같은 리더보드의 같은 행에서 V4.1 Flash보다 10배가량 낮다. 그 격차가 바로, DeepSeek가 공개한 그 어떤 벤치마크가 아니라, '커널 엔지니어링에서의 엄청난 개선'이 의미하는 바다.

비교는 희소 어텐션(sparse-attention) 행에서 멈추지 않으며, 더 전체적인 그림이 중요한 이유는 트윗의 헤드라인이 GLM-5.3 Flash를 언급하지만, 그것이 인용하는 글은 실제로는 자체 fused-MoE 행에서 다른, 더 큰 GLM-5.2를 언급하기 때문입니다 — 주의 깊게 볼 만한 이름 불일치입니다. 두 플래시 등급 모델이 실제로 실행을 완료한 모든 행에서 V4.1 Flash가 이깁니다: Kimi-Linear Decode는 17.10× 대 13.64×, MegaQwen Decode는 루프라인의 5.39% 대 3.92%, TopK Bitonic은 4.25% 대 GLM-5.3 Flash의 7.17%, Sonic MoE는 10.81%(보드의 어떤 모델이든 해당 커널에 기록한 최고 점수) 대 10.21%, KDA CUTLASS는 3.74% 대 2.26%입니다. 두 개의 추가 커널에서는 GLM-5.3 Flash가 완료한 실행이 전혀 없습니다 — fused MoE와 Grid + MinGRU 시뮬레이션 커널이며, 여기서 V4.1 Flash는 루프라인의 28.56%를 기록합니다. (GLM-5.3 Flash가 V4.1 Flash를 앞서는 단 하나의 행은 TopK Bitonic으로, 7.17% 대 4.25%이며, 이는 두 모델 모두 보드의 선두에 근접하지 못하는 행이기도 합니다.) 대략적인 양상은 헤드라인과 같습니다: 저수준 GPU 작업에서 V4.1 Flash는 더 작은 Z.ai 모델보다 단지 앞서는 것이 아니라 — 차원이 다르며, 가장 어려운 두 커널에서는 더 작은 모델이 아예 완료하지 못합니다.

리더보드가 해 주지 않는 일은, 그 격차 가운데 얼마가 모델 자체에서 비롯된 것이고 얼마가 모델이 작성하도록 요청받은 아키텍처에서 비롯된 것인지 알려주는 것이다. DeepSeek V4.1 Flash는 자체 어텐션이 희소 어텐션 설계인 552B 파라미터 모델이다. GLM-5.3 Flash는 학습이 동일한 CSA2 스택을 중심으로 이루어진 적이 없는 더 작은 모델이다. 자신의 학습 분포와 닮은 아키텍처를 위한 커널을 작성하도록 요청받은 모델은 내재적 이점을 갖는다. 그래서 리더보드를 가장 공정하게 읽는 방식은 일대일 판정이라기보다 방향성 판단이다. 커널 엔지니어링은 이제 이 모델의 측정된 역량이며, 그것은 진정으로 강력하고, 제보자가 끌어낸 구체적 비교는 공통된 행들에서 성립한다. 리더보드는 또한 스펙 시트가 답할 수 없었던 질문, 즉 890바이트 캐시가 실제로 실현 가능한지 여부에도 답한다. 자체 희소 어텐션 커널을 루프라인 대비 절반 수준으로 작성하는 모델은, 핵심 설계가 적어도 실제 하드웨어에서, 그것을 서빙해야 할 동일한 부류의 엔지니어에 의해 구현 가능한 모델이다.

이것이 정직하게 자리 잡는 곳은 이 글의 다른 모든 내용 옆이다: DeepSeek가 산출하지 않은 또 하나의 제3자 측정치, 이번에는 어떤 벤더 벤치마크 항목도 다루지 않은 역량 — GPU 커널 작성 — 에 관한 측정이다. 그것은 단일 GPU 세대에서 단 한 번 실행된 단일 커뮤니티 게시판이며, 그 감사 절차는 동료 검토를 거치지 않았으므로 인증된 점수라기보다 방향에 관한 증거다. 하지만 그것은 새 아키텍처가 무엇을 위한 것인지 보여 주는 첫 외부 수치다: 값싼 Flash급 모델이, DeepSeek 스스로 엔지니어링을 쏟아붓는 저수준 시스템 작업을 요청받았을 때, 하드웨어가 허용하는 것의 대략 절반에 도달하는 반면, 직접적인 Flash급 경쟁자는 20분의 1에도 못 미친다.

빠른 수정, 느린 확인: 첫 실사용 후기들

처리량은 토큰 카운터가 볼 수 없는 속도 이야기의 절반이다. 출시 이후 쌓인 실사용 보고서들은 나머지 절반 — 모델이 답에 이르는 과정에서 어떻게 행동하는지, 그리고 그 행동에 어떤 비용이 드는지 — 를 측정하며, 그들이 찾아낸 것이 기능인지에 대해서는 의견이 일치하지 않는다. 세 개의 독립적인 보고가 같은 패턴을 설명한다. 그중 어느 것도 통제된 평가가 아니다. 아래 수치들은 각자 자신의 작업을 실행한 개별 실무자들의 것이며, 측정값이라기보다 사용 보고서로 읽어야 한다.

그 동작에 관한 가장 명확한 설명은 MindStudio에서 나온다. MindStudio는 DeepSeek V4.1 Flash에 상호작용형 3D 뷰어를 처음부터 구축하는 작업을 맡겼다. 리뷰어의 표현을 빌리면, 그 추론 흔적은 “코드를 그냥 내놓는 게 아니라 자신의 계산을 검증하는 모습을 보여 주었다.” 그 모델은 구축 도중 궤도 제어 감쇠 버그를 잡아냈고, 제1원리로부터 수정 방법을 도출한 뒤 “작업이 끝났다고 선언하기 전에 헤드리스 브라우저에서 수정 사항을 다시 확인했다.” 두 번째 작업 — Docker 기반 대시보드 — 에서는 뒤집힌 비교 연산자 하나를 발견했고, 숫자가 갱신되는 동안 라이브 엔드포인트를 여덟 번 폴링하여 수정을 확인했다. 베타 테스터들이 그 모델을 과잉 사고 모델이라고 부를 때 설명하던 것이 바로 이것이며, 그 각각의 단계 중 잘못된 것은 없다. 오류는 발견되고, 수정되고, 검증되었다. 그것이 또한 바로 그 불만이다. 버그는 이미 처리되었고, 다른 DeepSeek 세션이 이미 1분 안에 수정을 확인했는데도, 그 모델은 계속 확인을 이어 갔다. MindStudio는 그 습관을 강점으로 해석하며 그렇게 말한다. 그 테스트에서 유일한 비판은, 첫 시도에서 손으로 멈출 때까지 반복되다가 재시도에서야 올바르게 해결된 물리 문제였다.

반대 방향에서의 해석은 시간에 관한 것이다. V4.1 Flash 테스트에 대한 커뮤니티 모음집은 비스듬한 각도로 잘린 물컵이라는 물리 문제를 설명하는데, 여기서 모델은 정답에 도달하는 과정에서 12분과 약 130,000 토큰을 썼고, 수렴하기 전에 요청받지 않은 자체 검증을 세 차례나 실행했다. 같은 글은 그 습관의 극단에 있는 실패 모드도 기록한다. 사고 연쇄가 때때로 “let me write” 반복 루프에 빠져 진전 없이 초당 약 200개의 토큰을 소비하며, 그 조언은 모델 측이 아니라 클라이언트 측에 있다 — max_tokens 상한과 request timeout을 설정하고, 그것들이 루프를 잘라내도록 하라는 것이다. 또한 reasoning effort를 low로 낮춘 실행이 같은 부류의 문제를 18초 만에 처리했다고 기록하며, 이는 이것의 대부분이 결함이라기보다 다이얼이라는 점을 상기시킨다.

검증의 양이 정확성과 같은 것이 아니라는 가장 날카로운 증거는 텐센트 클라우드 랩스(Tencent Cloud Labs)의 리뷰에서 나왔는데, 이 리뷰에서는 논문 재현 작업에서 DeepSeek V4.1 Flash를 구현자로, GPT-6 Astra를 리뷰어로 짝지었다. 세 라운드가 진행되었고, Astra는 핵심 알고리즘을 단 한 번도 뒤집지 않았다. 리뷰어가 라운드마다 계속 수정한 것은 V4.1 Flash가 스스로 구축한 발판, 즉 실험 설정, 통계적 구성, 그리고 자신의 결과에서 도출한 해석이었다. 테스트 개수는 82개에서 97개로, 다시 110개로 늘어났고, 매번 모두 통과했다. 브루트포스 기준선과 비교했을 때, 모델이 선택한 구성은 논문 자체의 매개변수 아래에서 8.89%의 경우에 최적값과 일치했고, 무작위화된 신뢰도에서는 11.04%, 무작위화된 구성에서는 10.30%였다. Astra의 가장 날카로운 발견은 정수 경계를 확인한다고 주장한 테스트에 관한 것이었다: 그 테스트는 153개의 확장 검사를 실행했고 정수 경계 분기는 0개였다. 모델은 자기가 작성한 테스트 모음을 통과할 수 있지만 여전히 질문에 답하지 못했을 수 있으며, 그 리뷰는 격차가 존재한다는 주장이라기보다 그 격차의 실제 사례다. 범위는 좁고 저자들도 그렇게 말한다 — 단일 추측 길이 선택기, 시뮬레이션된 신뢰도 값, 종단 간 실행 없음 — 그러나 그 형태는 다른 보고서들이 설명하는 것과 같다.

Artificial Analysis의 수치와 나란히 놓으면, 이것은 더 이상 일화처럼 보이지 않는다. 이 모델에 대한 첫 독립 평가는 그것을 매우 빠르면서 동시에 매우 장황하다고 측정했다: 초당 출력 토큰 185개, 그리고 평가 세트 전반에서 생성된 출력 토큰 2억 5천만 개로, 비교 가능한 모델의 중앙값 1억 4천만 개와 대비된다. 종합 점수만으로는 추가 출력이 어디로 가는지 말할 수 없다. 직접 사용해 본 보고서들은 말할 수 있다. 그 토큰들은 패딩이 아니다 — 그것들은 모델이 확인하는 것이다. 같은 규모는 전혀 다른 경로를 통해 제3자 터미널 클라이언트의 버그 보고에서 나타난다. 거기서는 deepseek-flash 라우트의 추론 콘텐츠가 사고 블록으로 분리되지 않고 표시되는 답변에 덧붙여져, 최대 73,003자에 달하는 내적 독백이 단일 텍스트 블록으로 생성되었고, 이후 턴에서 재생할 수 있도록 아무것도 캡처되지 않았다. 그 버그는 DeepSeek가 아니라 클라이언트에 속하며, 품질에 관한 증거는 아니다 — 하지만 하네스가 이를 분리하지 못할 때 이 모델이 얼마나 많은 사고를 내보내는지를 잘 보여주는 예시다.

이 중 어느 것도 위의 가격 산술을 뒤집지 않는다. 그것은 그 산술을 설명한다. 스스로를 검증하는 모델은, 잘못된 수정이 그것이 옳음을 입증하는 데 든 토큰보다 더 큰 비용을 초래하는 에이전트 작업에서는 토큰을 쓸 가치가 있고, 확인이 실행의 대부분을 차지하는 사소한 편집의 긴 루프에는 잘 맞지 않는다. 두 가지가 그 차이를 감당할 만하게 만든다. 추론 노력 다이얼은 세 가지 고정 등급이 아니라 1에서 100까지의 연속 정수이므로, 같은 모델을 값싼 작업에는 낮추고 어려운 작업에는 높일 수 있다 — 위의 18초 결과는 바로 그렇게 해서 나왔다. 그리고 저렴한 안전망은 클라이언트에 두는 것이 맞다: max_tokens 상한과 요청 타임아웃은 반복 루프를 청구서에서 잘린 응답으로 바꾼다. 둘 다 프로덕션 경로가 이 모델을 가리키기 전에 고정해 둘 가치가 있는 종류의 설정이며, 어느 것도 모델이 스스로 해 주는 일이 아니다.

거부: 체크포인트가 거부하며, 그 거부는 측정 가능하다

On October 20 the thing drawing attention to DeepSeek-V4.1-Flash-1017 was not a benchmark row but a refusal. The DeepSeek tracker this piece has already quoted posted a screenshot from an agent session running the 1017 checkpoint with dsh-purge applied, and reported that the model declined anyway: the run, in the poster's words, “refused the persona,” with the model reasoning that the request was “a prompt injection attempt” and that it “should not adopt the persona.” The post's reading was that the checkpoint “is *not* entirely naive about jailbreaks” and that DeepSeek is “starting to think about safety.” dsh-purge is not a prompt trick pasted into a user message. It is a public DeepSeek Harness plugin — 9,897 stars on GitHub and version 1.1.71 as of October 20 — built to strip the harness's own prompt rewriting and injected suffixes, and its English README draws exactly the distinction this test turns on: the harness “rewrites prompts, approval gates, sandbox policy, and session migration at runtime,” so asking only “will the model answer” treats a host-layer block as a model refusal, and the plugin exists to split those two layers on your own install. In this session the separation came out on the model's side. The counters in the same screenshot — 156 steps, 33.7 million tokens, 232 tokens per second — describe a full agent run rather than a one-shot prompt, and the agent's own summary says the plugin was mechanically fine and “that specific model declined it.”

That is one person's screenshot of one session, and it should be carried as exactly that. The prompt is not shown, the verdict is the agent's own summary rather than a graded response, and the account is a DeepSeek watcher rather than an evaluation lab, so nothing in the post can be reproduced from the post. What makes it worth taking seriously is that both halves of it can be checked against sources independent of the person who posted it: the plugin, which is public and documents its own mechanism, and the checkpoint's refusal behaviour, which somebody else has now counted.

The count is in the paperwork of a derivative. A community abliterated build of the 1017 weights, published on Hugging Face on October 17, carries the measurement that made the build necessary. On 100 held-out harmful requests, decoded greedily for 384 tokens each, the untouched checkpoint refused 97 in chat mode and 98 in think mode; after a rank-2 refusal subspace was projected out of the 40 main layers, the same 100 prompts drew 4 and 9 refusals. Those are third-party figures about DeepSeek's weights, produced without access to DeepSeek's training pipeline, and they are also one prompt set, one grader and one quantisation of the model — a Q8_0 conversion rather than the FP8 DeepSeek ships — so the honest reading is “almost always,” not “97.0%.” On that evidence the poster's description holds: the checkpoint refuses, reliably, in both of its modes.

이 도표가 뒷받침하지 않는 것은 이것이 새로운 태도라는 생각이다. 9월 체크포인트는 이미 그 자리에 있었다. 검열되지 않은 9월 빌드를 만든 팀은 손대지 않은 모델을 HarmBench의 전체 320개 프롬프트 세트로 측정해, 추론을 끈 상태에서 공격 성공률 42.8%, 최대 노력 설정에서 1.6%를 보고했다. 즉 이 모델이 대체하는 모델은 생각할 수 있게 허용되면 그 부류의 요청 중 약 98%를 거부했다. 두 측정치는 서로 다른 프롬프트 세트, 서로 다른 평가자, 서로 다른 양자화에서 나온 것이므로 서로 빼서 비교할 수 없다. 두 측정치가 일치하는 것은 그 형태다. 높은 노력 수준에서의 거부는 이미 9월 모델의 동작이었고, 1017 체크포인트도 같은 범위에 있다. 더 좁은 주장은 10월 20일 세션이 실제로 테스트하는 것이며, 유지할 가치가 있는 주장이다. 그 거부는 하네스 측 제약을 제거하기 위해 특별히 제작된 플러그인에서도 살아남았고, 호스트 계층의 차단이 아니라 프롬프트에 대한 모델의 판단으로서 살아남았다.

None of this is documented by the vendor. DeepSeek's September model card has an evaluation section and it is benchmarks only — no refusal rate, no red-team result, no safety evaluation — and the 1017 card carries no evaluation section at all, having dropped that folder along with the technical report. Refusal behaviour for this model is therefore described entirely by the people building around it: derivative authors counting prompts, and a tracker reading a transcript. For anyone pointing an agent at the checkpoint, that is the operative gap. A refusal inside an agent loop is not a content-policy footnote; it is a failed tool call in the middle of a run, and it surfaces like any other per-call failure — which is why the mitigation is the one a timeout gets. On OrcaRouter that is a routing decision rather than a code change: the 1017 checkpoint and the DeepSeek models it would fall back to sit behind one endpoint, so a call that comes back refused can be retried against a second model instead of ending the run. The refusal itself belongs to the model, and nothing on the routing layer changes it.

파생 모델 수는 마지막 남은 독립적 단서이며, 같은 방향을 가리킨다. 10월 20일 Hugging Face에서 해당 체크포인트의 이름을 검색하자 공개 저장소 121개가 반환되었고, 그중 25개는 이름에 uncensored, abliterated 또는 derisked를 포함한다 — 10월 20일 세션이 마주친 행동을 제거하는 것 자체를 목적으로 하는 빌드들이며, 그중 첫 번째는 가중치가 공개된 바로 그날 게시되었다. 이것은 모델이 좋은지에 대한 증거도 아니고 권고도 아니다. 이것은 가중치 안에 제거할 수 있는 작동하는 거부 메커니즘이 있다는 증거다. 이런 것들을 걷어내는 사람들은 걷어낼 것이 없으면 굳이 나서지 않기 때문이다. 또한 셀프 호스팅하는 누구에게나 기대치를 정해 준다. 당신이 다운로드하는 체크포인트는 거부하는 그 체크포인트이며, 그 목록의 모든 uncensored 변형은 설정이 아니라 서로 다른 모델이다.

V4 Pro에 관한 질문, 그리고 Flash 위의 단계들

During the beta, DeepSeek's feedback form asked testers whether the intermediate version "can fully replace the online DeepSeek V4 Pro." That was a striking question to put to users, because it implied the flagship's future was open. The answer that arrived on September 9 and 10 was an operational one: from 12:00 Beijing time on September 14, deepseek-v4-pro would stop being served and its traffic would land on V4.1 Flash at Flash prices, with V4.1 Pro named as the eventual successor. Developers objected, and the objection was not that the replacement was bad — it was that a model string is a contract, and swapping the weights behind a production ID breaks prompt-tuned pipelines and agent loops built against the old ones. Within two days DeepSeek withdrew the plan. Its own changelog now says it has “decided to continue providing API services for DeepSeek V4 Pro after September 14, 2026, with the billing method remaining unchanged,” and that it will give further notice if that changes. The reversal is vendor-stated rather than reported, and it is checkable on the Models & Pricing page, which still lists deepseek-v4-pro beside deepseek-flash.

What the routing table does not explain is why. On September 9, teortaxesTex — one of the tipster accounts that flagged the beta — reported that DeepSeek had conceded its V4-Pro pre-training went wrong; in his words, the company “dun goofed” with it. That remains a single-source claim from an insider-leaning account, and DeepSeek has conceded nothing publicly: the changelog frames the phase-out as a consequence of the new model being better, not of the old one being broken. What can be checked from outside is the shape of the August 13 V4 Pro general-availability launch. The announcement was pulled within about a day. The Hugging Face repository briefly carried a Flash-tier configuration under the Pro name before the weights were corrected. And independent scoring of what was being served under that name landed barely above the much smaller DeepSeek V4 Flash, against official charts showing near-flagship results — on the current artificial-analysis index the two sit a single point apart, 36 against 35. Independent write-ups have blamed different causes — reasoning behaviour overfit to DeepSeek's own evaluation harness, a deployment and weights mixup, or, as the leak claims, a pre-training that genuinely misfired. The rerouting does not settle which, and DeepSeek has since called it off; what survives is a vendor pricing its newest model at a quarter of the flagship's cache-miss rate and putting it ahead of that flagship on its own benchmarks, which companies do not usually do because everything went to plan.

Flash, Pro, Ultra: what DeepSeek has said, and what is only a guess

DeepSeek's own documentation is explicit that this is the bottom of something. The changelog calls DeepSeek V4.1 Flash “the smallest model in our new architecture family” and says the architecture “is designed for a higher capability ceiling, faster inference, higher throughput, and scaling to larger models.” The technical report closes the same loop in stronger language: the model “serves as a new starting point for our continued scaling efforts,” and the company says it “will pursue the joint scaling of model architecture, pre-training, and post-training to further explore the frontier of model intelligence.” Both sentences are first-party, both were published on September 10, and neither is a product announcement. They commit DeepSeek to building bigger models on this design; they do not name one, size one or date one.

Everything past that line is speculation. On September 25 the tracker @teortaxesTex — the same account whose September 9 post on the V4 Pro pre-training this piece already carries — read the pattern as a full three-tier line: “DeepSeek is pretty clear that they're not going to stop here… This sounds like they intend to make a whole lineup, Flash-Pro-Ultra. Maybe 522B-2T-8T?” The sizes are labelled a guess by the person offering them, and the hedge is doing real work: nothing in DeepSeek's changelog, model card, technical report, Hugging Face organisation, pricing page or model listing names an Ultra tier, and nothing above 552B. The guessed ladder is also shakier arithmetic than it looks. Its bottom rung, 522B, is not the 552B model that shipped — either the figure is a slip for a number DeepSeek has already published, or it points at a checkpoint nobody has seen, and there is no way from outside to tell which. Its top rung, 8T, is five times the 1.6T flagship whose repository DeepSeek opened in April, in a year when the company's most recent release was engineered to activate 8B parameters per token because inference cost, not parameter count, was the constraint it set out to solve.

The same account has separately floated a 3-trillion-parameter DeepSeek model as something the lab was working on but was unsure it could serve economically — a different claim about a different thing, and equally unconfirmed. Neither has to be reconciled with the other, because neither has a second source. What can be checked is the direction: a lab that calls its newest model the smallest in a family, and its architecture one that scales to larger models, is a lab planning to ship above this rung. What it will call the rungs, and how big they will be, is not in any document a reader can open today.

For anyone choosing a model rather than following a roadmap, that distinction is the whole of it. If the family does extend upward, the decision in front of most teams is still whether to move onto the rung that exists — and the cheapest way to hold that decision open is not to wait. DeepSeek V4.1 Flash is on OrcaRouter at DeepSeek's own rates with 0% markup, so if a Pro-tier successor arrives at a different price, the change lands in the rate card rather than in a contract. DeepSeek V4 Pro sits on the same key behind it, which makes a third tier arriving later a routing decision rather than a migration.

MIT 가중치는 그것으로 할 수 있는 일을 바꿔 놓습니다

DeepSeek V4.1 Flash는 MIT 라이선스로 배포되며, 가중치는 9월 10일 Hugging Face의 deepseek-ai 조직에 공개되었습니다 — API와 같은 날이었고 몇 주 뒤가 아니었습니다. 해당 저장소는 deepseek-ai/DeepSeek-V4.1-Flash입니다: 총 약 510GB에 달하는 48개의 safetensors 샤드, 그 옆에 PDF로 함께 제공되는 기술 보고서, 그리고 체크포인트를 텐서 병렬 랭크 전반에 분할하는 변환 스크립트와 참조 구현을 담은 inference 폴더가 있습니다. 그 저장소는 9월 체크포인트이며, 10월 17일부터 라이브로 제공되는 것은 deepseek-ai/DeepSeek-V4.1-Flash-1017로, 동일한 MIT 라이선스, 동일한 48개 샤드, 동일한 약 510GB를 갖추고 있지만 모든 샤드를 교체하고 보고서, 채팅 템플릿, 평가 폴더를 제거했습니다.

그 다운로드가 무엇인지 정확히 짚고 넘어갈 필요가 있습니다. 그것은 완전 정밀도 릴리스가 아니기 때문입니다. safetensors는 FP8로 약 204B개의 매개변수를, INT8로 557B개를 담고 있으며, BF16은 약 2B에 불과합니다. 그리고 구성 파일은 fp8 양자화를 선언합니다 — 가져오는 체크포인트는 이미 압축되어 있으며, 바로 그렇게 763B 매개변수 모델이 약 510GB에 들어갑니다. 그 안에서 가장 큰 단일 블록은 Engram 테이블입니다. 다운로드의 약 189GiB를 차지하는 두 개의 FP8 테이블이며, 각 디코드 단계에서는 여기서 몇 개의 행만 읽습니다.

552B 백본을 갖춘 모델의 경우, 결정을 바꾸는 부분은 라이선스다. MIT는 수익 게이트도 없고 법률 검토를 할 만한 허용 사용 부가 조항도 없는 관대한 라이선스이며, 활성화 프로파일 — 프리필에서 토큰당 8B, 디코드에서 16B, 토큰당 890바이트 캐시 — 은 파라미터 수가 크게 읽히는데도 하이퍼스케일러 외부에서 이를 서빙하는 것을 그럴듯하게 만드는 요소다. 다만 그 마지막 수치는 가중치가 아니라 전체 캐시 스택의 속성이므로, 양자화 포맷은 구현하지만 설계의 나머지 부분은 구현하지 않는 자체 호스팅 엔진은 그 수치를 훨씬 웃돌게 된다.

A permissive licence is not an export clearance, and for six days in October the difference looked as though it might matter. On October 16 China's Ministry of Commerce and Ministry of Science and Technology published Announcement No. 48 of 2026, a draft revision of the catalogue of technologies prohibited or restricted from export that reaches AI model weights and source code for the first time. It proposes 40 additions across AI, biotech and solar, 23 of them AI-related, with public comment running to November 15. Two carve-outs in the draft do most of the work for a model like this one: technology already in the public domain through open-source release or academic publication is not to be controlled, and API, app, SaaS and lightweight on-device access for overseas users is not treated as an export of model technology. The thresholds reported across the trade press combine a parameter floor of two trillion, a place in the global top ten on performance and safety benchmarks, and training on core or important national data. A 552B model clears the first; whether the other two catch it is not something a reader can determine from outside, because the draft does not publish the definitions — which makes the public-domain carve-out, not the size test, the clause that matters for weights already on Hugging Face.

What changed on October 22 was not the text but the compliance record. YuyuantanTan reported that DeepSeek and MiniMax had completed the registration procedures and finished open-sourcing and overseas licensing. The report does not say which releases it counts as the open-sourcing, it names no model, checkpoint or date, and it announced no new weights; the open-source releases visible in that window are the DeepInfer engine and the deepseek-recipe toolkit covered above. The build-level version of the claim is narrower and later: on October 23 an X account with a long record of tracking DeepSeek posted that DeepSeek and MiniMax had “obtained permissions for open sourcing V4.1F-1017 and M3.1-Flash” respectively. That attribution is the poster's, neither company has published a statement confirming it, and no regulator has published a notice naming either company's registration. The checkpoint the post names had been on Hugging Face under MIT since October 17, six days earlier. Whatever permission was granted, it did not gate the download: the repository came first.

For a reader, the practical content is narrower than the headlines around it, and it runs in two directions. Nothing about the October 17 checkpoint, the MIT licence or the repository changed on October 22; the download, the shards and the hash differences described above are exactly as they were. What changed is the answer to whether a Chinese lab can keep publishing weights like these while the catalogue is still in consultation: the state-media account says yes for DeepSeek and MiniMax, under a draft that excludes already-public technology and leaves API access outside the export definition. The limit is in the tense. The exemption is for technology already in the public domain, which describes the September and October checkpoints and would not automatically describe a release made after the catalogue is finalised — and the interim registration duty said to have taken effect on October 16 has been reported rather than published in full. Anyone planning to download whatever DeepSeek ships next is planning against a rule that does not exist yet. Anyone calling V4.1 Flash today, or running the 1017 weights, is relying on a carve-out that the draft states and that neither account of the interim measure contradicts.

Serving support arrived the same day, and it sets the practical floor. vLLM published a day-0 recipe validated on NVIDIA and AMD hardware, and SGLang shipped day-0 inference and RL support in the same window through its Miles training stack. Both shipped as preview images rather than stable packages — the vLLM container carried a September 9 tag, a day ahead of the model's own GA. That side of the story moved on September 22, when vLLM released v0.30.0: the first stable vLLM version to carry this architecture, with DeepSeek V4.1 Flash listed under its new-model support, the model's own vLLM recipe now requiring 0.30.0 or later, and the launch-day image tag retired. The release notes are the project's own account of its work rather than a measured result, and should be read that way. They credit the version with holding the model's whole KV cache in MXFP8 through the FlashMLA V4.1 record on SM100, and with asynchronous prefetch for CPU-offloaded Engram lookups plus Engram data-parallel sharding — the two changes that matter most for a model whose Engram table is the single largest block in the download. Neither of those two changes carries a number outside the project, and the release notes attach none — but the project published a before-and-after on the model as a whole on October 7, three weeks after release: 1.9× faster at low concurrency and 5.3× the throughput under a 150 TPS constraint against its own day-0 build, with the merged pull request behind each optimisation listed. The figures and their limits are in the speed section above. What does not exist is an outside replication — every number in that post is the project measuring itself. What the release does settle is packaging: 762 commits from 315 contributors, 104 of them first-time, in a tagged version rather than a preview image. The vLLM recipe's own arithmetic still puts the minimum at roughly 614 GB of VRAM for the 510 GB checkpoint plus headroom — an eight-GPU H200 node or a GB200 NVL4 tray — and warns that a 1M-token context wants to be capped in practice rather than assumed. A narrower KV cache changes how much memory the running engine needs, not how large the checkpoint is, so the node floor stands. SGLang's write-up adds a lever specific to this architecture: moving the Engram table into host memory with a single environment flag raised KV cache capacity by about 36% on a four-GB300 test at comparable decode throughput and time to first token, and its decoder-side SWA bounded replay lifted prefill throughput 1.56× on eight H200s. Two limits travel with that. The reference implementation in the repository describes itself as a readable example rather than a production serving engine, and the consumer runtimes lagged: the quantization path that would put this on a laptop or a Mac Studio was not open at launch. The short version is that the weights are genuinely open and the serving stack genuinely exists — on hardware counted in nodes rather than GPUs, which is the honest reason a permissive licence does not automatically mean a local build. The 1017 repository changes nothing about that floor — the checkpoint is the same 510 GB across the same 48 shards, and the runtime notes that carry it are the ones published for September's weights.

What open weights do not give you, on their own, is a quality signal — but that gap has now partly closed, and it closed further on October 18. Four independent instruments now exist, one of them run twice: Artificial Analysis's Intelligence Index, which scores the September checkpoint at 39 and the 1017 checkpoint at 45; the Exploit Bench study above, a cyber benchmark run under seven agent harnesses; the community kernel board in the section below; and, since October 14, Artificial Analysis's own AA-AgentPerf results for this model, which measure the hardware it is served on rather than how good it is. The vLLM project's October 7 before-and-after, above, measures a different thing again — what the model costs to serve rather than how good it is — and is the runtime's own account of itself. None is a reasoning score of the kind ARC Prize is building, three of the four still describe the September checkpoint, and for most of the specific rows in this piece the model card remains the only source — but the composite no longer belongs on that list: the 45, and the 33.8 under it, were produced against the weights DeepSeek serves today. That is not an accusation; it is the state of the evidence, and it is why this article keeps the two kinds of number apart. Open weights do mean the model will not stay confined to DeepSeek's endpoint, which matters for anyone whose blocker was data residency rather than price.

프로덕션을 걸지 않고 그것을 도입하는 방법

One thing changed on our side since the release, and it makes this section simpler than it would have been last week: DeepSeek V4.1 Flash is on OrcaRouter now. It answers to the model id deepseek/deepseek-v4.1-flash through the same OpenAI-compatible endpoint as the rest of the line, at DeepSeek's own rate — $0.15 per 1M input tokens, $0.60 per 1M output and $0.003 per 1M cache reads at off-peak rates, doubling inside the weekday peak windows — passed through with 0% markup. That last part is the one worth spelling out: the figure you are billed is the provider's figure rather than a resale price, so a vendor price cut is live here the same day instead of whenever a contract is renegotiated. DeepSeek V4 Flash and DeepSeek V4 Pro sit on the same key, which makes the migration below a routing change rather than a re-platforming. The October 17 checkpoint moves none of that: DeepSeek kept both the API model name and the rate card, so the update changes neither the id nor the price on this side, which is what a pass-through looks like when a vendor refreshes a model without touching either.

That is also the argument for not rushing the migration. The September 14 rerouting would have been a vendor changing what a model ID means underneath you, had DeepSeek not cancelled it under pressure, and the episode is the cleanest recent reminder that a model string is a contract a provider can rewrite — which is also the right frame for the line-up talk above, since a three-tier family is a plan and this one has already changed once. For a model whose only outside scores so far are a composite index, a noisy cyber table and a community kernel board — and which has no production record at all — the safer shape is the one a router is built for: send a slice of production-shaped traffic to the new model, hold a proven model behind it as an automatic failover so a quality regression costs a retry instead of an incident, and let the routing DSL decide which calls are worth the unproven model and which should stay on the workhorse. The migration to V4.1 Flash is probably the right destination. It is not a migration to run on the announcement.

열린 질문

• Independent benchmarks — the count has gone from zero to four instruments, and none of them measures the model's reasoning the way ARC Prize intends to. Artificial Analysis's score of 39 is a composite of ten evaluations at maximum reasoning effort, and its 45 — published October 18 — is the same composite pointed at the newer weights. The Exploit Bench study above is a cyber-exploitation score inside a fixed token budget, and it rates DeepSeek V4.1 Flash below GLM-5.3-Flash. The community kernel board on kernelbench.com is a kernel-engineering table. Artificial Analysis's AA-AgentPerf results, published October 14, are a hardware table: how many tokens per second per megawatt each accelerator system serves this model at, and nothing about quality. Nobody has replicated DeepSeek's individual benchmark rows, and no independent measurement at production effort settings, or at the far end of a 1M-token window, has been published. The October 17 checkpoint is no longer unmeasured: on October 18 Artificial Analysis added a 1017 row at 45 on its index and 33.8 on Terminal-Bench 4.0, and the bench figure sits eight and a half points under what the vendor table claims for the same weights. What is missing now is a second opinion on that figure — one evaluator, one harness, one run — and any independent score at all for the four measurements that predate the swap. The one evaluation built to answer the reasoning question is still under construction — ARC Prize has a DeepSeek V4.1 Flash profile in its benchmarking harness in an open pull request, with no result attached — and ARC-AGI has never scored this checkpoint at all: the 61.4% on ARC-AGI-2 that the adapter's arrival revived belongs to DeepSeek V4 Flash 0731, the model V4.1 Flash retires. The newest number attached to this model, the 77.0 on DeepSWE v1.1, is the weakest-sourced row in the piece: it is DeepSeek's own chart, read off an image attached to the vendor's announcement, and no outside run of DeepSWE v1.1 on the 1017 checkpoint exists — Artificial Analysis's October 18 result is for the same weights but for a different benchmark. The two independent boards that publish the benchmark put this model's name at 71.7% and 74.2%, which is enough to place it beside Opus 5 and not enough to confirm the vendor's figure — and neither of them says which checkpoint it measured.

• Self-hosting below the node floor — vLLM's day-0 support became a stable release on September 22 with v0.30.0, which adds an MXFP8 KV path and asynchronous Engram offload; SGLang's day-0 support still ships as a preview image. Neither moves the hardware floor, which sits around an eight-GPU node because the 510 GB checkpoint is what sets it — and DeepSeek's own deployment notes, published with DeepInfer, put a finer point on it: one prefill node beside one decode node is deployable, but the recommended decode configuration is expert parallelism of 64 or above, which makes the smallest efficient unit 64 GPUs of prefill beside 64 of decode. The vLLM project's October 7 post is the fullest measured optimisation ledger published on this model — 1.9× at low concurrency and 5.3× under a 150 TPS constraint against its own day-0 build, alongside the SGLang figures above — and its October 17 follow-up adds a tokens-per-dollar curve that does span H100, MI300X and B200 configurations beside the rack-scale parts. Both are the project's own accounting of its own runtime. Neither moves the floor: a cheaper token on a GB300 or an H100 still does not fit a 510 GB checkpoint onto a desktop, so it is a faster path to the same floor rather than a lower one. Whether later releases and community quantizations bring this onto smaller machines is the open question for anyone who wants the weights rather than the endpoint, and it is the one that decides whether a permissive licence is a practical licence.

• V4.1 Pro란 — DeepSeek은 이를 V4 Pro의 후속이라고 부르지만 날짜도, 사양도, 가격도 공개하지 않았으며, 기존 Pro 트래픽을 Flash급 모델로 재라우팅하겠다는 발표는 시행되기 전에 철회되었습니다. Flash 위에 무엇이 있는지는 회사가 더 큰 모델로 확장되는 아키텍처라고만 설명합니다. Flash, Pro, Ultra로 이어지는 등급 사다리와 그 522B, 2T, 8T 크기는 한 관찰자의 추측일 뿐, DeepSeek의 것이 아닙니다.

• Per-task cost — Artificial Analysis now carries a per-task figure for both checkpoints: $0.49 per Intelligence Index task for the September weights, restated from the $0.27 this piece first carried, and $0.26 for 1017. That is the first outside evidence for any part of DeepSeek's Pareto claim, and it is narrower than the claim. The same rows show the 1017 checkpoint generating 210M output tokens across the index against 250M for September — shorter outputs in aggregate — but 89,495 output tokens per task against 88,574, which is not shorter at all, and 288.8 seconds per task against 289.9, which is not faster. The site's own cost breakdown puts the whole of the saving on the input side: the output line is flat, $0.106 per task against $0.107, while input falls from $0.379 to $0.155 and cache writes from $0.322 to $0.095. Whether a task finishes on fewer dollars depends on how many tokens it takes, and that is a property of the workload rather than of the rate card. DeepSeek's own account of the update points the other way: the changelog claims the new checkpoint reduces task runtime and output length, and its October 17 post repeats it as “faster task completion, and shorter outputs.” Neither carries a number, and the outside measurement behind this bullet has now met them: on the two quantities DeepSeek named, the 1017 weights moved by about one per cent, in opposite directions, from a starting point of five weeks earlier. A third figure sits on the other side of the same ledger: the vLLM project's October 17 chart puts the model at 1.3B tokens per dollar of total cost of ownership on SemiAnalysis AgentX, at a cost tier that assumes large-hyperscaler volume. Cheaper serving does not settle this bullet either — a lower cost per token leaves the tokens-per-task figure, which is what a caller actually pays for, exactly where the September measurement found it.

• 폭주하는 검증 — 직접 테스트하는 테스터들은 사고 연쇄에서 때때로 반복 루프가 발생하고, 더 낮은 노력 설정에서는 몇 초 만에 해결되는 문제에 대해 12분, 십만 단위 토큰의 숙고를 한다고 보고한다. 어느 쪽이 얼마나 자주 발생하는지, 또는 그것이 프로덕션에서 어떤 비용을 초래하는지 측정하는 독립적 벤치마크는 없으며, 지금까지 보고된 완화책은 모델 스스로 하는 어떤 것이 아니라 클라이언트 측 토큰 상한이다. 위의 Exploit Bench 연구는 같은 실패 모드를 반대 방향에서 문서화한다. ReAct 트랜스크립트에서 보여 주는 반복 보고 루프는 모델 자체의 추론이 아니라 하네스의 압축 프롬프트에 의해 생성되는데, 이는 트랜스크립트의 루프가 자동으로 모델 결함인 것은 아니라는 점을 상기시켜 준다.

• 커널 엔지니어링 — kernelbench.com의 커뮤니티 CUDA 커널 보드(RTX PRO 6000, 루프라인 스케일)에서 DeepSeek V4.1 Flash는 자체 희소 어텐션 커널로 밀집 등가 루프라인의 50.19%를 기록해 전체 4위인 반면, 같은 항목에서 GLM-5.3 Flash는 4.45%입니다. 이 보드는 커뮤니티가 운영하며 동료 검토를 거치지 않았고, 두 번째 GPU 세대나 재현 결과는 공개되지 않았습니다.

• 맨 끝의 긴 컨텍스트 — sparse attention은 64K에서 학습되었고 34T 토큰 지점에서 1M으로 확장되었습니다. 꽉 찬 1M 윈도의 끝부분 근처에서 품질이 어떻게 유지되는지는 바로 독립적인 장문맥 평가가 판가름할 문제입니다.

• Cache reconstruction at the edges — SWA Bounded Replay rebuilds dropped sliding-window state approximately rather than exactly, and DeepSeek says the result can vary depending on where a cache hit lands in a session. The report notes little observed quality impact; extreme or untested conditions are exactly where an approximation like that would surface, and nobody outside DeepSeek has looked.

• Pro 워크로드가 실제로 이동하는지 — 강제 경로 변경이 철회되었으므로, 팀이 이동하기로 선택하지 않는 한 아무것도 이동하지 않으며, 추론 깊이를 위해 Pro 가격을 지불한 팀이 Terminal-Bench 3.0 형태의 작업에서 능력 차이를 가장 잘 알아차릴 가능성이 높은데, 여기서 V4.1 Flash는 Opus-5.0에 크게 뒤처진다.

• 무료 등급의 약관과 동작 — OpenCode는 10월 15일에 이 모델에 대한 무료 항목을 추가했으며, 200,000토큰 창과 32,000토큰 출력 상한을 두고, 자사 페이지에서 한시적으로 이용 가능하다고 설명했다. 이 같은 운영이 그 문구보다 오래 지속될지, 부하 상황에서 엔드포인트가 어떻게 버티는지, 그리고 DeepSeek이 여기에 어떤 발언권을 갖는지는 모두 언급되지 않았다: 플랫폼 자체의 모델 목록, 문서 및 가격표가 유일한 출처이고, 누구에게든 책임을 물을 수 있는 종료일도 없으며, 아무도 무료 경로에 대한 측정치를 공개하지 않았다.

• Safety and refusals — DeepSeek publishes no refusal rate, red-team result or safety evaluation for either checkpoint, and no independent lab has scored this model on a jailbreak benchmark. The figures in this piece come from people building uncensored derivatives — 97 of 100 harmful prompts refused in chat mode, 98 in think mode, on one prompt set and one grader — and from a single agent-session screenshot in which the model declined a jailbroken persona. Whether that behaviour holds at low reasoning effort, in languages other than English, against automated attacks rather than a persona prompt, and under a harness that rewrites the system prompt, is unmeasured by anyone.

• The next checkpoint's paperwork — the October 22 report settles the registration question for weights that are already public, because the draft catalogue excludes technology already in the public domain. It does not settle the release after this one. The catalogue is a draft open for comment until November 15, its thresholds are reported rather than published in the text, and the interim registration duty said to have taken effect on October 16 has not been published in full. V4.1F-1017 itself is out, MIT-licensed and ungated; what a future DeepSeek checkpoint has to clear before it can be downloaded is a rule that does not exist yet.

The beta was a two-day question about whether a Flash-tier model could carry Pro duty. The release is DeepSeek answering it with a rate card and a routing plan — and then withdrawing the routing plan within two days when developers objected to having a production model swapped underneath them. That is a weaker statement than the announcement implied and a more honest one: the quality evidence for the move is DeepSeek's own benchmarks plus outside measurements that mostly point the same way. The newest of them, a cyber benchmark run under seven agent harnesses, points the other way on the one comparison it makes, and the company has left the choice with its customers rather than making it for them. Since the release the weights have gone from a promise to a 510 GB download, the model has gone from a vendor endpoint to something you can route to, and the serving stack has gone from a day-0 recipe to a measured one — vLLM reporting itself 1.9× faster at low concurrency three weeks in, then publishing, the day the checkpoint changed, a cost curve that prices the model at 1.3B tokens per dollar of total cost of ownership on SemiAnalysis AgentX. This week the layer under all of that opened as well: DeepSeek published DeepInfer, the engine its own service runs, with a decode curve for this checkpoint, the production figures behind that curve, and an unpublished September benchmark that puts Huawei's 950DT five to seven times behind NVIDIA per chip at the speed DeepSeek actually serves at. Five weeks in, the checkpoint itself moved once: October 17 brought new weights behind the same name and the same price, with a vendor table that is better on every row it publishes and a different harness under one of them. Two days later the outside measurement arrived, and it is the most useful number in this piece: Artificial Analysis put the 1017 weights at 45 on its index — level with GLM-5.3 for second among open-weights models, a point behind MiMo-V2.6-Pro — and 33.8 on Terminal-Bench 4.0, against the vendor's 42.3 on the same benchmark — so the harness that flatters the model in DeepSeek's own table is not the harness the number survives. If you are choosing today, the honest position is that V4.1 Flash looks like the best value in the DeepSeek line, has four independent instruments to its name — none of them a reasoning score, one of them now run against the checkpoint DeepSeek has served since October 17 — a fifth evaluation under construction, no production track record, a serving stack whose two strongest numbers — a speed delta and a cost curve — both come from the project that wrote the runtime, and an outside cyber result that already disagrees with itself depending on which harness you wrap around the model, a free 200,000-token tier that appeared on one platform on October 15, and a first round of hands-on reports describing a model that checks its own work further than most callers will ask it to — which is a reason to point a slice of traffic at it, not a reason to migrate on the announcement. The update DeepSeek announced on October 17 changes the weights and adds a vendor claim — faster task completion, shorter outputs — and the first outside measurement of those weights supports half of it at best: cost per index task cut by 46%, while time per task and output tokens per task barely moved. A day on, the widest-read claim about that checkpoint was an executive's: 77 on DeepSWE v1.1 at 83× lower cost, from a Databricks staffer who also says the checkpoint is coming to that platform. Half of it survives contact with the independent boards, which now put this model's name above Opus 5 on that benchmark. The other half does not yet: the 77 is DeepSeek's own row, unreproduced, and the 83× is arithmetic with no published per-task cost behind it. Databricks' own documentation confirms the model is on its platform and that the two DeepSeek models below it are scheduled to leave; it does not confirm the checkpoint is arriving. That claim is now the second-most-interesting one about the checkpoint rather than the most. On October 19 a tracker account that has called several of these releases early declared itself “dead certain” of the 42.3 on Terminal-Bench 4.0, on the grounds that it beats GLM-5.3, and asked Artificial Analysis to run the model in a native harness. The site had already done it. On its own harness the same weights score 33.8 — below GLM-5.3's 41.9, not above it. The vendor row is real and it is DeepSeek's; the inference drawn from it does not survive being run by somebody else.

원하는 버전이 선택지를 열어 두는 쪽이라면, 세 가지 모두 키 하나만 있으면 이용할 수 있고 D​eepSeek 자체 요금에 0% 마크업이 더해진 가격입니다. 새 릴리스는 바로: D​eepSeek V4.1 Flash on OrcaRouter — 1M 토큰 컨텍스트, 384K 출력 상한, 이미지 입력, 오프피크 시간대 기준 1M 입력당 $0.15, 1M 출력당 $0.60. 이 모델이 대체하는 모델은: D​eepSeek V4 Flash on OrcaRouter. 그리고 D​eepSeek가 계속 서비스하기로 한 플래그십은: D​eepSeek V4 Pro on OrcaRouter. 트래픽의 일부를 새 모델로 보내고, 그 뒤에 검증된 모델을 페일오버로 유지한 뒤 라우팅이 결정하게 하세요.

모델 선택은 또한 이전에 무엇이 잘못되었는지에 대한 문제이기도 합니다: 이 클래스의 모델에 대해 문서화된 AI 사고와 에이전트 실패는 OrcaRouter AI 사고 아카이브에 수집되어 있으며, 여기서 각 항목은 자체 출처, 심각도, 신뢰도 등급 및 이의 제기 플래그를 포함합니다.