DeepSeek-V4.1-Flash-1017: A 45-to-46 Guess, a Three-Sentence Changelog, and the One Row Both Scoreboards Share

D​eepSeek-V4.1-Flash-1017 is real and it is one day old. You can read the changelog entry dated 2026-10-17, download 763 billion parameters of MIT-licensed weights from Hugging Face, and call it right now under the unchanged deepseek-flash model string at the unchanged rate card. What you cannot do is look up where it ranks. The most substantive post about DeepSeek V4.1 Flash the day after its update landed was not a result — it was a guess. A close follower of the lab put the model's eventual Artificial Analysis Intelligence Index score at 45 to 46, predicted Grok 4.7 as the number it nudges past and AA-Omniscience as the evaluation that flattens it, and offered 57% as a fair expectation on the Vals Index. Every one of those figures is a forecast about a scoreboard that has no 1017 row on it yet. The release is a fact. The rank is a bet, and the interesting part is why nobody can settle it today.

What actually shipped on October 17

The primary source is short. D​eepSeek's changelog entry for 2026-10-17 says V4.1 Flash "has been upgraded to D​eepSeek-V4.1-Flash-1017," that the web and app models moved with it, that "the API calling method and pricing remain unchanged," and that calls to deepseek-flash reach the newest model. That is the whole specification-level claim. The corresponding repository, deepseek-ai/D​eepSeek-V4.1-Flash-1017, was created at 02:40 UTC on October 17 and last modified at 05:46 UTC the same day; its safetensors index totals 763,205,315,794 parameters across BF16, F32, F8_E4M3 and I8 tensor types. The licence is MIT. Nothing about it is a leak.

The changelog adds three things that are not numbers. It calls the update a "performance–cost Pareto improvement." It names the harnesses it was tuned against — D​eepSeek Harness, T​encent CodeBuddy and WorkBuddy — and tells users to move to D​eepSeek Harness 0.2.1-rc.1 or later. And it shows four project demos, rendered as animation: a 3D model of the Forbidden City, a browser racing game, a Portal 2 remake built in Godot, and a 3D scene called "Valley of Ages." Those are marketing artifacts, not measurements, and they are the reason this piece keeps them in a separate paragraph from the scores.

What did not change is as load-bearing as what did, because it is the part with money attached:

• Rate card — unchanged. Cache-hit input $0.006 per million tokens at peak and $0.003 off-peak; cache-miss input $0.30 / $0.15; output $1.20 / $0.60. Peak means 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday, excluding Chinese public holidays.

• Model string — unchanged. deepseek-flash now resolves to the 1017 checkpoint. The retired names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and still route to it, billed at Flash rates.

• Envelope — unchanged at the card level. 1M-token context window, max_tokens at 256K or above, temperature 1.0 and top_p 1.0 recommended.

• Packaging wart — also unchanged. There is still no Jinja chat template. The repository ships encoding/encoding.py as a reference implementation covering multi-turn conversation, tool calling, thinking mode, numeric reasoning effort, mid-conversation system messages and interleaved images, and points production users at deepseek-recipe, a Rust library with Python bindings.

The 1017 card does not restate the architecture. The family's shape — 552B backbone, a Causal Encoder-Decoder split into a 20-layer causal encoder and a 20-layer decoder, asymmetric activation of 8B parameters during prefill and 16B during decode, 384 routed experts with 6 active per token, CSA2 attention, FP4 main-KV caching, and a 196B Engram conditional memory — comes from the September card, the technical report titled Pushing the Limits of KV Cache Compression, and the model card's own citation. A reader who wants those numbers should take them from the September document, not from a 1017 page that never claims them.

Eleven vendor rows, and the one the index actually reads

The vendor benchmark table is the only comparable artifact in the release, and it is published as an image on the changelog and as a markdown table on the model card. Read from the card, the 1017 column against the September column: HLE without tools 42.6 (44.0 on the text-only subset) against 36.8 (39.1); 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; ZeroBench-main with tools 51.0 against 49.0. All eleven rows move up. Two footnotes qualify the table: the dagger symbol marks the text-only subset of HLE, and Terminal-Bench 4.0 was run in the C​laude Code harness while DeepSWE v1.1 was run in Mini-SWE-V2.

Now hold that against the Artificial Analysis Intelligence Index, which is the thing the 45-to-46 guess is about. The current index version is 4.3.2, and it is a weighted composite of ten evaluations under four headings: Agents at 30%, Coding at 20%, Scientific Reasoning at 20%, General at 30%. The weights are published, and they are the whole argument. AA-Briefcase v1.1 carries 15%. AA-Omniscience carries 15% split into accuracy at 10% and a non-hallucination rate at 5%. Terminal-Bench 4.0 carries 10%. HLE carries 10%. CritPt, the research-level unpublished-physics set, carries 10%. GDPval-AA v2.1 carries 10%, GDP.pdf 10%, SciCode 10%, AutomationBench-AA 5% and AA-LCR v1.1 5%.

Only one of those ten evaluations appears in D​eepSeek's eleven-row table under the same name and the same configuration: Terminal-Bench 4.0, at a tenth of the index. HLE appears on both sides but not identically — the vendor column is the pure-text subset, and the index runs the 2,158-question text-only set. Everything else the index reads, including the two heaviest components, has no vendor row at all. That is the mechanics behind the guess, not a criticism of the lab: it is simply that a vendor table and a composite index are built from different parts, and the overlap is one row wide.

Where 45–46 comes from, and why you cannot check it

The forecast is arithmetic on the September score. Artificial Analysis currently scores the September DeepSeek V4.1 Flash checkpoint at 39 in its reasoning configuration at maximum effort, which places it ninth of 120 models on the Intelligence Index and well above the median of 19 for open-weight models of its size class. The same entry carries $0.27 cost per index task, 175.9 output tokens per second and 1.23 seconds to first token. The 45-to-46 forecast is therefore a claim of roughly six or seven index points from one checkpoint swap — an unusually large single-step move, which is exactly why the poster hedged it with "would be decent" rather than predicting it.

Landing there would put 1017 in crowded company. The models currently at 46 are Grok 4.7 at xhigh and high, MiMo-V2.6-Pro and G​PT-6 Astra at low effort. At 45 sit Qwen3.8-Max (0902), M​use Spark 1.3 at xhigh and G​LM-5.3 at max. At 44 sit S​tep 5 Preview and Kimi K3 at max. Six or seven points would move D​eepSeek's cheap tier from below the open-weight cluster into the middle of it, and the comparison the poster chose — "better than G​rok" — is the one that a 45 or 46 would make literally true at Grok 4.7's own effort setting, while costing an order of magnitude less per task. That is the useful shape of the guess: not the integer, but the price-to-score position it implies.

There is one more reason to be careful with the baseline itself. A third-party write-up published in September quoted the September checkpoint at 40 on the index, at a rank of sixth of 113; the site reads 39 at ninth of 120 today, on a named index version. Same weights, different published integers, because the index is versioned and its population grows. Any forecast built on a number that has already moved twice is a forecast built on sand, and the honest version of the 45–46 claim is a range about a range.

The Omniscience prediction is the one with a mechanism

Of everything in the signal, the claim that 1017 "will get tanked by Omniscience again" is the part with a testable mechanism behind it rather than a vibe. AA-Omniscience is not a knowledge quiz with a pass mark. It scores factual reliability across roughly 6,000 questions spanning 42 topics, awards points for correct answers, deducts for hallucinations and leaves abstentions neutral — so declining to answer beats guessing wrong. A model tuned for agentic throughput and short outputs, which is what the 1017 changelog describes, is a model with fewer tokens of deliberation per answer. Whether that helps or hurts on an index that penalises confident wrong answers is a real open question, and "again" is doing work in that sentence: it implies the September checkpoint already underwhelmed there. The index publishes Omniscience accuracy and non-hallucination rate as separate numbers, but the D​eepSeek model page does not surface per-benchmark figures, so the September Omniscience split is not currently readable from the site. The prediction is plausible and, today, unfalsifiable.

The other half of the same complaint — that the index "still uses broken CritPt" — points at a genuine property of the composite. CritPt is research-level unpublished physics: problems whose answers are not yet in the training data, which is what makes it resistant to contamination and also what makes it noisy and slow to refresh. It carries a tenth of the index, the same weight as HLE and Terminal-Bench 4.0. A model family that has never published a CritPt row is carrying a 10% blind spot into every forecast made about it. Both halves of the criticism are really the same observation: the index and the vendor are measuring different things, and the forecast has to bridge that gap with assumptions.

The Vals number is a different board with a different base

The 57% figure deserves its own paragraph because it is the one part of the signal that maps onto a published number, and the mapping is a year out of date in the right direction. Vals AI's index places the September DeepSeek V4.1 Flash at 57.86%, fifteenth of 56 models and the top open-weight entry, marginally ahead of Kimi K3 at 57.81%, at a stated $0.30 per test. So "57% on Vals sounds fair" is not a projection upward — it is a projection of roughly flat performance on a board where the September checkpoint was already the open-weight leader. Two boards, two different expectations, from the same poster, in the same sentence: a big step up on the index and no step at all on Vals. That asymmetry is the most informative thing in the post, because it tells you the poster expects the 1017 gains to land in the evaluations the index weights heavily and that Vals weights lightly — Terminal-Bench-style agentic work, not general knowledge-work breadth.

What would actually have to be re-run

None of the independent numbers in circulation describe the 1017 weights. D​eepSeek replaced the model behind a stable API name, which means the ownership of every published figure has quietly changed hands: a score quoted for DeepSeek V4.1 Flash from a September run now describes weights the vendor no longer serves, while the weights it does serve have no external score at all. For anyone who treats a checkpoint swap as a configuration change, the practical rule is the one drawn from software: it warrants the same regression pass. Five questions are worth answering against 1017 before it goes near a production path.

• Does Terminal-Bench 4.0 reproduce? The vendor moved it 31.2 to 42.3 in the C​laude Code harness, and it is one of the few rows an independent board reads. A vendor harness choice that flatters a model is the oldest failure mode in benchmark reporting.

• Does the DeepSWE gain survive a different scaffold? 77.0 in Mini-SWE-V2 against the September 74.2. DeepSWE has shown large spread across harnesses, so the delta is the claim, not the level.

• Does output length actually fall? The changelog's cost claim rests on shorter outputs, not just faster ones. On a metered API that is the difference between a Pareto improvement and a rounding error.

• Does the reasoning-effort dial behave? Effort is numeric with low, high and max mapping to defined values, high is the default, and temperature is accepted but ignored in thinking mode. Test at the exact effort value production sends, because a suite run at max tells you nothing about a deployment at high.

• Does multi-turn tool use still round-trip? With tools, reasoning_content from every earlier turn — including turns with no tool call — has to come back, or the request fails. That is the kind of regression a checkpoint swap introduces silently.

If you want to run that regression pass without moving a production path onto unmeasured weights, this is the case OrcaRouter's failover is built for: put 1017 behind a fallback chain, send it the traffic you are willing to lose, and let a failing or slow response retry across providers before the caller sees it. Pricing is the other half of the argument and the simpler one — D​eepSeek kept the Flash rate card unchanged, and because OrcaRouter passes the provider's list price straight through with 0% markup, a change that does not move D​eepSeek's price does not move ours. You can reach DeepSeek V4.1 Flash at D​eepSeek's own rates today, with no second contract and no code change, while the leaderboard catches up.

The questions the signal leaves open

Why would a checkpoint swap move the index at all, if the architecture is the same? Because the index weights agentic and coding evaluations at half its total, and those are the evaluations where post-training changes of this kind show up fastest. The September model's weakest published category was agentic work relative to its price; the 1017 changelog is almost entirely about agentic work and cost per task. The forecast is a bet that the same architecture plus better post-training lands where the weights say it should.

Is there any way to read a 1017 number before Artificial Analysis publishes one? Only indirectly. The two agentic rows the vendor publishes — Terminal-Bench 4.0 and Automation-Bench — are the closest proxies, and both are self-reported. A third path is to watch whether the September entry on an independent board is renamed, re-run or left alone; a stale row that keeps its old score while the served weights change is the clearest signal that a number no longer means what it says.

Does the unchanged price mean the Pareto claim is real? It means the claim is testable. "Performance–cost Pareto improvement" at a fixed rate card is only true if output tokens per task fall, because that is the only lever left. The vendor says task runtime and output length both dropped. That is the single claim on the page a reader can verify with their own bill.

What to watch next

The forecast to grade is specific, which is what makes it worth keeping. If a 1017 row appears on the Intelligence Index at 45 or 46, the poster called the position and the mechanism — an agentic-heavy weighting plus a known soft spot on Omniscience — and the useful lesson is that the composite became predictable from the vendor table once the weights were read carefully. If it lands at or below the September 39, the lesson is the opposite and more expensive: an eleven-row vendor table that moves up on every line can still leave a composite score flat, because the rows that moved are not the rows that count. Either outcome is one poll of one model, not a verdict on the lab, and the number that decides it is one a reader cannot compute today.

What a reader can do today is cheaper than a forecast. The weights are MIT-licensed and the API is live at the September price, so the month before anyone publishes a rank is the month to form your own view on the three questions that matter to you: does it finish your agentic tasks faster, does it say less while doing so, and does it hold up on the multi-turn tool paths you actually run.

Source note: every D​eepSeek figure in this piece — the eleven benchmark rows, the 763B parameter total, the 1M-token window, the MIT licence, the unchanged rate card and the harness recommendations — comes from the vendor's own October 17 changelog entry, its model card on Hugging Face, and its published pricing page. Every independent figure — the 39 on the Intelligence Index, the ninth-of-120 placement, the $0.27 cost per index task, the 175.9 tokens per second, and the ten-evaluation weighting that carries the score — comes from Artificial Analysis, read on October 18, 2026. The 45-to-46 forecast, the 57% Vals expectation and the Omniscience prediction are one observer's estimates about a model that has no 1017 row on any of these boards, and they are labelled as estimates everywhere they appear.

The documented AI failures and agent incidents behind releases like this one are collected in the OrcaRouter AI Incident Archive, where each entry carries its own sources, severity, confidence grade and disputed flag.