A generated title card reading 'LingSante-3.1-flash vs Qwen 3.8', subtitled 'One reads the note, one reads the scan', above two rounded panels divided by a 'vs' disc. The left panel is headed LingSante-3.1-flash with a pale blue 'MIT licence' badge and four rows: 560B total, 25B active; text in, text out; 1,000,000-token context; self-hosted only. The right panel is headed Qwen3.8-27B with a pale amber 'Apache 2.0' badge and four rows: 27B dense; text, image, video in; 262,144-token context; 0.33 and 2.40 per 1M. A footer reads 'LingSante-3.1-flash figures are vendor-reported and unannounced; Qwen3.8-27B 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: One Reads the Note, One Reads the Scan

Auteur

jinhao.song

Date de publication

Derniers modèles · 20Voir tous les modèles →
Benchmarks : Artificial Analysis · mis à jour quotidiennement
Retour à tous les articles

The boldest sentence on the LingSante-3.1-flash model card is that it "ranks second overall behind only GPT-6 Astra" on benchmarks for medical professional behaviour. That claim deserves a paragraph of its own, and it gets one below. But it is not the decision in front of most teams, because the model you would actually compare this one against is not a frontier lab's flagship — it is Qwen3.8-27B, the Apache-licensed generalist that runs on a single high-end workstation, reads documents and images, and costs $0.33 per million input tokens on OrcaRouter. A 560-billion-parameter medical specialist that appeared three days ago against a 27-billion-parameter generalist that half the industry has been shipping since August: that is a real choice, and it turns on things the marketing sentence does not mention. The general half of this decision is a priced call today, not a procurement project. the rest of the hosted catalogue Every model on it is billed at the provider's list price, with no markup added.

The first of them is blunt. LingSante-3.1-flash takes text. Qwen3.8-27B has a vision encoder. In a hospital, the input is very often a scan of a form, a PDF of a discharge summary with a lab table rendered as an image, or a photograph. One of these two models can be pointed at that file directly.

The claim, and the rows under the claim

Four rows on the LingSante card measure something like professional medical conduct, and reading them individually is more informative than reading the summary sentence.

• PhysicianBench — LingSante-3.1-flash 57, second in the table behind Gemini 4 Argon's 58 and ahead of Claude Opus 5.5's 53.

• HealthBench Professional — 65.72, first in the table, above GPT-6 Astra's 64.18 and Claude Opus 5.5's 61.19. This is the strongest single number the model has.

• MedSCE, a safety benchmark — 83.11, third, behind Tencent Hy4 at 86.53 and DeepSeek-V4.1-Flash at 84.22.

• MedEthiAlign, an ethics benchmark — 80.85, fifth of nine, behind Tencent Hy4 at 88.83, Claude Opus 5.5 at 88.08, GLM-5.3 at 85.84 and Gemini 4 Argon at 82.71.

First once, second once, third once, fifth once. No single row in the table places LingSante-3.1-flash second behind GPT-6 Astra specifically, so the headline sentence is a composite of the vendor's own choosing, computed on a weighting the card does not publish. That does not make it false. It makes it unverifiable, which is a different thing, and it is the right frame for everything else on the card: these are vendor runs, on a harness the vendor selected, summarised by a rule the vendor kept private.

Worth saying plainly, because it is easy to lose in the table: the underlying numbers are genuinely good. A 65.72 on HealthBench Professional is the best figure in a comparison that includes four frontier models. The scepticism is about the summary, not the rows.

Screenshot of the Hugging Face model page for inclusionAI/LingSante-3.1-flash captured 22 November 2026 in English: the header with 1 like, the tags 'bailing_hybrid', 'custom_code', 'fp8' and 'License: mit', a panel reading 'Model size 572B params' with tensor types BF16, F32 and F8_E4M3, an 'Inference Providers' box stating 'This model isn't deployed by any Inference Provider', and the card's introduction stating 560 billion total parameters, 25 billion active and a 1-million-token context window, with the italic line that the model is intended for research and is not a medical device.

Two rows overlap, and neither is medical

Qwen3.8-27B's card and LingSante-3.1-flash's card have exactly two benchmarks in common. Neither is a medical benchmark, because Qw​en's card publishes no medical benchmark at all and LingSante's publishes no GPQA.

• SWE-bench Pro — LingSante-3.1-flash 65.25 against Qwen3.8-27B 61.7. LingSante's card notes its run used the Claude Code v2.1.205 harness; Qw​en's card does not name a harness. Same benchmark, possibly not the same conditions.

• HLE, text only — LingSante-3.1-flash 44.07 without tools against Qwen3.8-27B 30.8. Qw​en's card does not state whether tools were available, so treat the 13-point gap as indicative rather than measured.

There is a trap next to these two, and it is worth naming because comparison sites walk into it constantly. LingSante-3.1-flash reports Terminal-Bench 4.0 at 26.5. Qwen3.8-27B reports Terminal Bench 2.1 at 73.0. Different versions of a benchmark that was substantially rebuilt between them. Putting those two numbers in the same sentence produces a 47-point gap that means nothing. The honest reading is that LingSante's card does not report a terminal-agent result comparable to Qw​en's, and the 26.5 is best read as "this model was not built for shell work".

So the overlap is thin, and what it shows is unsurprising: on general reasoning the medical specialist is ahead, on coding-agent work it is behind, and on medicine nobody has measured Qwen3.8-27B at all. If your question is "which is better at medicine", the published evidence answers it only in one direction, because only one side published.

A medical model that cannot read a scan

This is the part of the comparison that survives all the benchmark noise. LingSante-3.1-flash inherits the Ling-3.1-flash architecture, which is a text model: 560 billion parameters, 25 billion active, a million-token context, and no vision tower anywhere in the config. Qwen3.8-27B is a causal language model with a vision encoder, and its card reports OmniDocBench 1.5 at 91.1 for document intelligence, CharXiv at 90.2 with confidence intervals, and MathVision at 94.6.

Clinical text does not arrive as clean text. It arrives as a faxed referral, a scanned consent form, a lab report exported as an image, a PDF whose tables were never tagged. A text-only model needs an OCR step in front of it, and that step is where transcription errors enter — a dropped decimal in a potassium value is exactly the kind of failure that no amount of downstream reasoning recovers from. A model that ingests the page directly removes one system from the pipeline and one class of error with it.

None of this makes LingSante-3.1-flash the wrong tool. A million-token window over text records is a different and real capability, and a model trained on medical text with an evidence-reasoning objective is not interchangeable with a general model reading the same note. But if your ingestion path is documents and images, the specialist cannot start the job.

A million tokens, natively and otherwise

Both cards advertise a million-token window, and they get there differently. Qwen3.8-27B states 262,144 tokens natively and up to 1,000,000 with YaRN, which is an explicit admission that the far end of the window is an extrapolation. LingSante-3.1-flash states 1,000,000 tokens and does not say whether that figure is native or extended; the Ling-3.1-flash family it is built on has been described elsewhere as extending to a million by the same method.

The distinction matters because a window is not a retrieval system. A million tokens of patient history is only useful if the model finds the right four sentences inside it, and the evidence for that is one row. LingSante-3.1-flash reports AA-LCR at 84.3, which is a long-context reasoning score and a good one. Qwen3.8-27B's card publishes no long-context benchmark at all — no AA-LCR, no NoLiMa, no RULER. So the specialist has a measurable long-context claim and the generalist has an unmeasured one, which is the reverse of the situation on document ingestion and worth holding in mind before either number is quoted in a procurement document.

What a case note costs on each side

Take a 40,000-token case file and a 1,500-token answer, and price both routes.

• Qwen3.8-27B through OrcaRouter — $0.33 per million input tokens and $2.40 per million output, so about $0.0132 for the file and $0.0036 for the answer. Call it 1.7 cents a case, with 262,144 tokens of native context and a million available with YaRN.

• LingSante-3.1-flash — no hosted price, because there is no hosted endpoint. The weights are 573.8 GB in fp8 and want 8 × H200; the MXFP4 build that landed today needs roughly half that hardware. At an assumed $2 per GPU-hour, eight H200s cost $16 an hour, so the specialist breaks even against the hosted generalist at about 950 cases an hour.

A generated two-column scoreboard titled 'LingSante-3.1-flash vs Qwen3.8-27B - the scoreboard' with the same six row labels on both sides. Left column LingSante-3.1-flash reads: Parameters 560B total, 25B active; Input text only; Context 1,000,000 tokens; HealthBench Pro 65.72 vendor; SWE-bench Pro 65.25; Price self-hosted only. Right column Qwen3.8-27B reads: Parameters 27B dense; Input text, image, video; Context 262,144 native; HealthBench Pro not published; SWE-bench Pro 61.7; Price 0.33 / 2.40 per 1M. A footer reads 'LingSante-3.1-flash figures vendor-reported; Qwen3.8-27B price from the OrcaRouter catalogue, benchmarks vendor-reported.' The OrcaRouter logo is composited in a white strip along the bottom edge.

That is the whole economic argument in two lines. Below roughly a thousand cases an hour, the hosted 27B is cheaper than owning the 560B. Above it, the fixed cost amortises and the specialist's marginal cost per case falls to electricity. Most organisations are nowhere near the crossover, which is why the practical question is not "which model is better" but "how much of this workload is medical reasoning versus document handling" — and document handling is the part Qwen3.8-27B does natively.

If you want the true size-class fight rather than the practical one, it exists: Qwen3.8-2.4T-A95B, 2.4 trillion total parameters with 95 billion active, whose weights Alibaba opened alongside the 27B. That is a different purchase with a different hardware bill, and it is not the model most teams are choosing between.

Where the router earns its place

One of these two models is a call. Qwen3.8-27B is in the OrcaRouter catalogue at the list price above, with no markup added on top, so a provider price change reaches your bill the same day rather than at the next contract renewal. The other is a cluster. LingSante-3.1-flash is not in our catalogue and will not be while it is a 294-to-574 GB checkpoint; it runs behind your own endpoint or not at all.

Screenshot of the OrcaRouter model page for qwen/qwen3.8-27b captured 22 November 2026 in English: the model heading 'Qwen3.8 27B' with Vision, Tools, JSON and Reasoning tags, a spec panel reading 262K tokens context, input text + image + video, output text, p50 TTFT 2.96 s and zero data retention available, a price strip showing $0.33 per 1M input tokens and $2.40 per 1M output tokens with p50 TTFT 2.96 s and 199.8M tokens of traffic over 7 days, and a Python code sample calling the model through the OpenAI-compatible endpoint at api.orcarouter.ai/v1.

That split is exactly what automatic failover is for. A pipeline that sends document ingestion and general reasoning to a hosted model and only escalates the genuinely medical judgement to a self-hosted specialist has one dependency that can be re-routed without notice and one that cannot. Putting the re-routable half behind a single key with failover across providers is the cheap half of the resilience problem, and it is the half most teams skip.

Qui devrait choisir lequel ?

If your input is text and your question is medical reasoning at volume — prior authorisation letters, evidence synthesis, structured extraction from narrative notes — LingSante-3.1-flash is worth the evaluation, provided you have the four GPUs the MXFP4 build needs and can accept that every medical number you are evaluating against came from the vendor.

If your input is documents, images or scans, or your workload is mixed enough that a single model has to do ingestion and reasoning, Qwen3.8-27B is the model you can start with this afternoon, and its vendor-published OmniDocBench and CharXiv figures are the reason.

If you have neither a GPU budget nor a document problem, neither of these is your answer — a smaller general model on your own hardware will do the job, and the comparison that matters is a different article.

What would change this verdict is specific and near-term: an independent medical evaluation of LingSante-3.1-flash, since Artificial Analysis has no page for it today; a hosted endpoint, which would collapse the cost argument into an ordinary per-token comparison; or a reproduction of the MXFP4 build showing what the four-bit repack costs in accuracy. Until one of those lands, the defensible position is that the 560B model has the better medical numbers and the 27B model has the better deployment story, and only one of those two things can be tested by a reader of this article.

Comparés dans cet article2

Détecté à partir de cet article · Benchmarks : Artificial Analysis · mis à jour quotidiennement