A generated title card reading 'LingSante-3.1-flash vs Qwen3.8 Flash' with the subtitle 'Metered tokens or a GPU bill'. Two rounded panels sit either side of a 'vs' disc: the left is headed LingSante-3.1-flash with rows '560B total, 25B active', 'Text only', 'MIT weights' and 'Your GPUs or a plan'; the right is headed Qwen3.8 Flash with rows '125B total, 6B active', 'Text, image, video in', '$0.15 in / $0.47 out' and 'Metered per token'. A footer reads 'LingSante figures vendor-reported; Qwen price from the OrcaRouter catalogue.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Guides & Insights

LingSante-3.1-flash vs Qwen 3.8 Flash: Metered Tokens or a GPU Bill

Tác giả

jinhao.song

Ngày đăng

Mô hình mới nhất · 20Xem tất cả mô hình →
Benchmark: Artificial Analysis · cập nhật hằng ngày
Quay lại tất cả bài viết

Ten thousand patient cases a day at 40,000 tokens each is 400 million input tokens. Priced at the OrcaRouter catalogue rate for Qwen3.8 Flash, that input costs $60 a day before a single token comes back. Run the same volume through LingSante-3.1-flash and the question changes shape entirely, because there is no meter to read: you buy GPUs or you buy a subscription, and the marginal case is close to free. Both models have "flash" in the name. Only one of them is priced like a flash model, and that is the fork in the road this comparison is really about.

The second fork is what each one is for. Qwen3.8 Flash is a general multimodal model with about six billion active parameters, built to move a very large volume of text, images and video through a pipeline cheaply. LingSante-3.1-flash is a 560-billion-parameter medical specialist with 25 billion active, built to reason carefully about a small number of high-stakes cases. They overlap on one thing — both can read a long clinical document — and diverge on almost everything else.

Two models, one adjective

Qwen3.8 Flash shipped in late August 2026 as the production build of the Qwen3.8-Flash-Next open-weight preview. It is a mixture-of-experts model with 125 billion total parameters, a 51-billion-parameter N-gram embedding table, and roughly 6 billion parameters active per token. It reads text, images and video, holds 262,144 tokens natively and reaches a million through YaRN, which is how Alibaba deploys it by default. On the OrcaRouter catalogue it runs at $0.15 per million input tokens and $0.47 per million output, with cached input at $0.0184.

LingSante-3.1-flash surfaced as weights on 19 November 2026 and was announced officially by Ant Ling on 22 November, alongside an MXFP4 quantised build. It is also a mixture-of-experts system, but a much larger one: 560 billion total parameters, 25 billion active, 63 Kimi Delta Attention layers and 9 gated MLA layers, 512 routed experts with 8 activated per token. It is text-only. Its card claims a million-token context; the shipped configuration is 262,144. The licence is MIT.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 23 November 2026 in English: the header with 27 likes and the tags Text Generation, Safetensors, bailing_hybrid, conversational, custom_code, fp8 and License: mit; the introduction stating 560 billion total parameters, 25 billion activated per token and a 1-million-token context window; the italic research-and-evaluation notice; and the medical benchmark chart. The sidebar shows 'Model size 572B params' with tensor types BF16, F32 and F8_E4M3, and an Inference Providers box stating 'This model isn't deployed by any Inference Provider'.

The price arithmetic, worked

Start with the case. A 40,000-token input and a 2,000-token answer, which is a long history plus labs plus a guideline excerpt, with a short reasoned response.

• Qwen3.8 Flash, no caching — 40,000 tokens at $0.15 per million is $0.0060. 2,000 tokens at $0.47 per million is $0.00094. Total $0.0069 per case, or about $69 a day at ten thousand cases.

• Qwen3.8 Flash, cached guideline — the 20,000-token guideline block that does not change between cases moves to the $0.0184 cache rate. That block falls from $0.0030 to $0.00037, and the case lands near $0.0043. Call it $43 a day.

• LingSante-3.1-flash, subscription — Ant Ling has not published a per-token rate. Its launch materials point to a Computrix token plan whose basic tier is quoted at US$20 a month, with no stated allowance, which means you cannot compute a per-case cost from it at all.

• LingSante-3.1-flash, self-hosted — the MXFP4 build is 291.72 GiB and fits on four H20-3e GPUs; the FP8 build is roughly 578 GB across eight H200s. At a spot rate near $2 per GPU-hour, a four-card node costs about $8 an hour, or roughly $5,800 for a month of continuous service. That figure moves with region and commitment, but the order of magnitude does not.

Put those together and the crossover is easy to see. At ten thousand cases a day, Qwen3.8 Flash costs about $1,300 to $2,100 a month depending on cache hit rate. A dedicated four-GPU node costs about $5,800 a month and serves only this model, with capacity you may not fill. Below that volume the metered model wins on cost without argument. The specialist's case is not price. It is that the node is yours, the data never leaves it, and the marginal case really is free once the node is paid for.

Qwen3.8 Flash sits on the OrcaRouter catalogue at exactly the numbers above, because provider list pricing is passed through at 0 percent markup rather than marked up with a platform fee. Alibaba's own price changes therefore land on your bill the day they happen, with no renegotiation. LingSante-3.1-flash is not on our catalogue; it runs through the vendor's plan or on your own SGLang deployment, and the model card recommends tensor and expert parallelism across eight H200s.

Six billion active versus twenty-five billion

Price follows compute, and the active-parameter count is where compute comes from. Qwen3.8 Flash activates about 6 billion parameters per token; LingSante-3.1-flash activates about 25 billion. That is a little over four times the arithmetic per generated token before you account for attention, which is why one of these can be sold at fifteen cents per million input tokens and the other cannot.

The throughput claims follow the same logic. Qwen reports up to eight times faster decoding on long cached contexts than its predecessors, and positions Flash as roughly five times faster than Qwen3.8-Max. LingSante-3.1-flash publishes no speed figure at all. Its card gives sampling settings — temperature 1.0, top_p 0.95, top_k 20, thinking enabled — and a hardware recipe, which is what you do when the serving target is a datacentre you do not control.

For an intake pipeline this is the whole story. A model that activates six billion parameters and accepts images can clear a queue of a hundred thousand documents overnight on rented capacity. A model that activates twenty-five billion parameters and takes text cannot, and was never meant to.

What Qwen3.8 Flash reads that LingSante cannot

The modality difference is absolute. Qwen3.8 Flash takes text, image and video in; LingSante-3.1-flash takes text. If the input is a photograph of a medication box, a faxed lab sheet, a scanned discharge summary or a short video of a gait assessment, the generalist can be pointed at it and the specialist cannot be pointed at it at all.

That capability is measured, not just asserted. Qwen's card reports 95.7 on MathVision with a code interpreter, 88.5 on RealWorldQA, 76.6 on LVBench and 84.5 on AndroidWorld, alongside 91.7 on GPQA Diamond, 91.9 on LiveCodeBench v6 and 62.5 on SWE-bench Pro. Those are vendor runs on Qwen's own harness, and they describe a model that is genuinely competent at reading a screen or a chart, not one that merely accepts an image token.

The honest counterpoint is that none of those numbers is a medical evaluation. A model that scores 95.7 on MathVision has demonstrated that it can extract numbers from a chart image; it has not demonstrated that it can tell a normal ECG from an abnormal one, and nobody has published a result that says it can.

The rows you would expect the generalist to win

Here is the surprise in this matchup. On the two general benchmarks both vendors happen to publish, the medical specialist is ahead.

• Humanity's Last Exam, text-only, no tools — LingSante-3.1-flash reports 44.07; Qwen3.8 Flash reports 35.9.

• SWE-bench Pro — LingSante-3.1-flash reports 65.25; Qwen3.8 Flash reports 62.5.

Both are vendor figures, run on harnesses each vendor chose, so the eight-point HLE gap should be treated as a signal rather than a fact. But the direction is consistent with what these models are: LingSante-3.1-flash is built on a 560-billion-parameter general base with 25 billion active, and Qwen3.8 Flash deliberately activates a quarter of that to hit its price point. A larger active set buys reasoning depth; the medical training did not remove it.

The same caution applies to the rows that look like comparisons but are not. LingSante's 72.22 on Humanity's Last Exam is with tools, not without, and its 94.10 on BrowseComp is a with-context figure; neither belongs in the same column as a no-tools run. Terminal-Bench versions differ between cards, so a Terminal-Bench 4.0 score of 26.5 for LingSante-3.1-flash cannot be set against any Terminal-Bench 2.x number at all.

The medical rows nobody can check

LingSante-3.1-flash's actual case rests on benchmarks only it publishes. HealthBench Professional is 65.72, above GPT-6 Astra at 64.18 and Claude Opus 5.5 at 61.19 in the vendor's table. MedQA-USMLE is 94.50, MedXpertQA-Text 57.27, EHR-Complex 74.40, PhysicianBench 57 and MedSCE, a safety benchmark, 83.11. The card also claims second overall behind GPT-6 Astra on medical professional behaviour, a composite computed on a weighting it does not disclose.

None of this has been reproduced. As of the announcement, the only independent figure on record is an Artificial Analysis MLCR of 23.3 percent, cited by the BenchLeader tracker, which lists no benchmark index for the model because it lacks the minimum number of independent results. The card itself says the model is for research and evaluation, is not a medical device, and may produce outputs that are incorrect, incomplete or unsupported by the evidence they cite.

Qwen3.8 Flash publishes no medical benchmarks, so there is no row on which to compare the two directly. That asymmetry is the thing to hold on to: the specialist's advantage is asserted on a surface where the generalist has made no claim, and the generalist's advantage is measured on a surface where the specialist has made no claim. Anyone presenting a clean winner is inventing a comparison that the evidence does not support.

Routing both behind one key

Screenshot of the OrcaRouter model page for qwen/qwen3.8-flash captured 23 November 2026 in English: the Qwen3.8 Flash header marked FEATURED and ZDR, dated 2026-08-26, tags for Vision, Tools, JSON and Reasoning, a specification panel listing 1M tokens context, 131K max output, input text + image + video, output text, p50 TTFT 10.00 seconds and a zero-data-retention note; a pricing row showing $0.15 per 1M input tokens, $0.47 per 1M output tokens and 4374.1M tokens of traffic over 7 days; plus a Python snippet pointing at api.orcarouter.ai/v1.

The architecture that follows from all of this is a two-stage pipeline. Qwen3.8 Flash takes the front — document handling, extraction, classification, the volume work — because it is cheap, fast and sees images. LingSante-3.1-flash takes the back — the differential, the evidence check, the summary a clinician reads — because that is what it was trained for. The two are not substitutes, and the cost model makes the split natural: the stage that runs ten thousand times a day is metered and inexpensive, and the stage that runs two hundred times a day is on hardware you already own.

Running that through one endpoint is where the routing layer earns its place. Qwen3.8 Flash is a single call on the OrcaRouter catalogue, the same key as every other model, with automatic failover if a provider degrades mid-pipeline. The routing DSL lets you pin the extraction stage to the cheap model and reserve the expensive one for the cases that clear a confidence threshold, without building two client stacks. Model fusion is worth a mention for the review step: a panel answering together and disagreeing is exactly the signal you want when the alternative is a single confident wrong answer.

What none of that fixes is the specialist's absence from the catalogue. LingSante-3.1-flash is not something we route, and a pipeline that needs it has to own the serving problem — the four H20-3e node, the SGLang configuration, the on-call rotation. That is a different kind of commitment than adding a model name to a config file, and it should be priced as one.

A split that holds up at volume

Use Qwen3.8 Flash when volume, latency and image input decide the job: intake, extraction, document triage, anything that has to clear a queue. Its $0.0069-a-case cost with no caching, falling to about $0.0043 with a cached guideline, is the reason it exists, and its 91.7 on GPQA Diamond says it is not merely cheap.

Use LingSante-3.1-flash when the output will be read by someone who will act on it, the input is text, and the volume is low enough that a GPU node or a subscription makes sense. Its medical scores are the best available in the open-weight field on the vendor's own evidence, and that evidence is the weakest link in the recommendation.

The thing to watch is the first independent medical evaluation. Until it lands, every advantage LingSante-3.1-flash claims is a claim, and the sensible deployment is one where a human reads the output and the model's refusals are logged. Qwen3.8 Flash does not have that problem — not because it is better, but because it is not claiming to be a medical model, and there is nothing to verify.

A generated two-column scoreboard titled 'LingSante-3.1-flash vs Qwen3.8 Flash - the numbers'. Both columns carry the same six labels. LingSante-3.1-flash reads 'Active params: 25B', 'Input: text only', 'HLE no tools: 44.07', 'Context shipped: 262,144', 'GPQA Diamond: not published' and 'Price: no rate card'. Qwen3.8 Flash reads 'Active params: 6B', 'Input: text, image, video', 'HLE no tools: 35.9', 'Context shipped: 1M via YaRN', 'GPQA Diamond: 91.7' and 'Price: $0.15 / $0.47'. A footer reads 'Both vendors unaudited; Qwen price from the OrcaRouter catalogue.' The OrcaRouter logo is composited in a white strip along the bottom edge.

So sánh trong bài viết này2

Phát hiện từ bài viết này · Benchmark: Artificial Analysis · cập nhật hằng ngày