A generated title card reading 'LingSante-3.1-air-VL vs LingSante-3.1-flash' with the subtitle 'Two medical models created 36 seconds apart'. Three icon cards read '124B vision vs 560B text', 'Six shared benchmark rows' and 'Flash wins all six'. A footer reads 'Vendor-reported scores; no independent run of either model.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Guides & Insights

LingSante-3.1-air-VL vs LingSante-3.1-Flash: Flash Wins Every Shared Row

ผู้เขียน

jinhao.song

วันที่เผยแพร่

โมเดลล่าสุด · 20ดูโมเดลทั้งหมด →
เบนช์มาร์ก: Artificial Analysis · อัปเดตทุกวัน
กลับไปยังโพสต์ทั้งหมด

At 04:30:30 UTC on 19 November 2026, inclusionAI created a Hugging Face repository for LingSante-3.1-flash. Thirty-six seconds later it created one for LingSante-3.1-air-VL. Same laboratory, same day, same medical brief, same MIT licence, and neither model announced anywhere. They are not two sizes of one thing. One is a 560-billion-parameter text reasoner built on Ling-3.1-flash; the other is a 124-billion-parameter vision model built on Ling-3.0-flash-VL, with 5.5 billion parameters active per token. Only one of them accepts an image, and only one of them is convincing when the reasoning has to happen without one.

The tempting comparison — "which is the better medical model" — has no answer, because the two cards were built to answer different questions. But there is a harder, more useful comparison hiding in them: six medical-text benchmark rows appear on both scorecards, and on all six LingSante-3.1-flash wins. That result is worth taking seriously precisely because it is so one-sided, and because it tells you what the smaller model's 5.5B active budget was spent on.

Same lab, different shift

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 23 November 2026 in English: 23 likes, tags Text Generation, Safetensors, bailing_hybrid, conversational, custom_code, fp8 and License mit; links to Hugging Face, ModelScope and a Computrix API; sidebar showing 'Model size 572B params' with BF16, F32 and F8_E4M3 tensor types, the notice 'This model isn't deployed by any Inference Provider', and the LingSante collection with 6 items updated about 13 hours ago; introduction stating 560 billion total parameters, 25 billion per token and a 1-million-token context window; a research-and-evaluation-only notice; and the top of the embedded benchmark chart.

The names suggest a family with two trim levels. The repositories describe two different machines.

• Backbone — LingSante-3.1-flash is built on Ling-3.1-flash, the 560B general model open-weighted in October. LingSante-3.1-air-VL is built on Ling-3.0-flash-VL, the September vision model. That is a generation apart, and the architectures reflect it.

• Layers and attention — flash is a hybrid-linear stack: 63 K​imi Delta Attention linear layers and 9 gated MLA layers in a 7:1 ratio, with four dense layers. air-VL has 42 layers, the first two dense, with multi-head latent attention and a 512-rank KV compression throughout.

• Experts — both route to 512 experts with eight activated per token. flash adds one shared expert and uses a 4,096 hidden size; air-VL uses 2,560 with no shared expert.

• Active parameters — 25B per token for flash, 5.5B for air-VL. Both are sparse; the difference is the size of the working set.

• Vision — air-VL carries a 27-layer ViT with a 4,096-dimension projector into the language model. flash has no vision path at all; its pipeline tag is text generation.

• Footprint — 578 GB of FP8 weights for flash, or 291.72 GiB in the MXFP4 build published on 22 November. air-VL is 249.7 GB of bfloat16 in fifty shards.

So the family hierarchy is not "big and small". It is "text reasoner" and "vision specialist", with different base models, different attention stacks and different serving images.

The six rows where they meet

A generated two-column scoreboard titled 'Six shared rows, six wins'. The left column, LingSante-3.1-flash, reads 'MedQA-USMLE: 94.50', 'MedMCQA: 83.89', 'DiagnosisArena-MCQ: 83.90', 'ReDisQA: 92.65', 'MedXpertQA-Text: 57.27' and 'HealthBench Professional: 65.72'. The right column, LingSante-3.1-air-VL, reads 'MedQA-USMLE: 94.34', 'MedMCQA: 81.54', 'DiagnosisArena-MCQ: 81.70', 'ReDisQA: 90.81', 'MedXpertQA-Text: 50.65' and 'HealthBench Professional: 49.59'. A footer reads 'Vendor-reported; both model cards published 19 November 2026, unaudited.' The OrcaRouter logo is composited in a white strip along the bottom edge.

Both cards publish a medical-text block, and six rows line up exactly. All figures are vendor-reported from the embedded tables; neither model has an independent run. Flash first, air-VL second.

• MedQA-USMLE — 94.50 vs 94.34. A 0.16-point gap, which is noise.

• MedMCQA — 83.89 vs 81.54. Flash ahead by 2.35.

• DiagnosisArena-MCQ — 83.90 vs 81.70. Flash ahead by 2.20.

• ReDisQA — 92.65 vs 90.81. Flash ahead by 1.84.

• MedXpertQA-Text — 57.27 vs 50.65. Flash ahead by 6.62.

• HealthBench Professional — 65.72 vs 49.59. Flash ahead by 16.13.

Six rows, six wins. Two are close enough to call a tie in practice; the last two are not close at all.

Flash wins all six — here's the gap

The size of each gap tracks how much clinical knowledge the question demands. MedQA and MedMCQA are multiple-choice recall; a 5.5B-active budget can clear those. MedXpertQA-Text and HealthBench Professional are different animals: they ask for reasoning over messy, realistic case material, and there the gap widens to seven and sixteen points.

That is the signature of capacity spent elsewhere. air-VL's active budget goes into visual tokens — a 27-layer encoder, 2,304 position embeddings, a 4,096-dimension projection — and there is less left for the kind of long-form clinical argument HealthBench rewards. The model can describe what is in a radiograph and loses the argument about what to do next.

On flash's own card that 65.72 is first in its row outright, ahead of GPT-6 Astra at 64.18 and Claude Opus 5.5 at 61.19. A medical fine-tune beating the frontier models on the benchmark closest to clinical judgment is a striking claim, and it is exactly the kind that needs an outside run before anyone builds on it.

What air-VL measures that flash cannot

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-air-VL captured 23 November 2026 in English: tags Image-Text-To-Text, Safetensors, bailing_moe_v3_vl, conversational, custom_code and License mit; sidebar showing 'Model size 125B params' in BF16, 'This model isn't deployed by any Inference Provider', and the LingSante collection; introduction stating 124B parameters, 5.5B active parameters per token and a 256K-token context window; a research-only notice; the full embedded benchmark chart with three rows of bar panels including AFUMED-Derm, AFUMED-Report, CheXpert Plus, MIMIC-CXR, MedXpertQA-MM, Medical VQA, DiagnosisArena-MCQ, HealthBench Professional, OmniDocBench v1.5, BrowseComp-VL, MMSearch and AndroidWorld; and Key Capabilities bullets describing imaging reports, 120-plus skin conditions, medical search and GUI agents.

The comparison above is deliberately unfair to the vision model, because it only counts rows where text is the input. air-VL's card runs twenty-six rows; flash's runs twenty. Six overlap. The other twenty on the air-VL card are tests flash cannot take: chest radiographs, pathology slides, medical VQA, document parsing, visual search and GUI agents.

On those, air-VL is strong on its own scorecard. First place on CheXpert Plus at 46.21, MIMIC-CXR at 47.20, IU X-RAY at 64.34, OmniMedVQA at 93.64, VQA-RAD at 81.15 and OmniDocBench v1.5 at 91.20. Second on PathVQA, PMC-VQA and SLAKE. The comparator set includes Claude Opus 5.5 and GPT-6.1 Sol at maximum effort, so these are not wins against straw men — they are wins against frontier models, asserted by the vendor, unverified.

The two cards also use the same AFUMED prefix for different tests, which is a trap for anyone skimming. air-VL reports AFUMED-Derm at 70.01, AFUMED-Food at 72.80 and AFUMED-Report at 90.70. flash reports AFUMED-LCR at 85.00 and AFUMED-Drug at 91.64. Same in-house family, different tasks, no overlap. A "vs" table that lines these up is comparing a skin-condition set to a medication-safety set.

Context: 256K documented, 1M claimed

The context story is the clearest place where the newer-looking model is the less finished one — and it runs opposite to the naming.

LingSante-3.1-flash promises a one-million-token window in its introduction and specification table. The configuration that ships with the weights sets max_position_embeddings to 262,144, and the MXFP4 sibling describes a "configured 262,144-token context window" without repeating the million. The quickstart publishes no rope-scaling block, so the million-token mode exists in the vendor's evaluation harness but not in a recipe a reader can copy.

LingSante-3.1-air-VL claims 256K and documents it. The base maximum is 131,072; the launch command passes a YaRN block with factor 2.0 and a flag permitting a longer context override. Whatever else is uncertain about a four-day-old repository, the 262,144 figure is reproducible from the card.

For a clinical workload that difference matters more than it sounds. A patient record, a prior study report and a guideline excerpt can push past 32K comfortably. Knowing that the long-context configuration is tested and published is worth more than a larger number with no recipe behind it.

The GPU bill separates them more than scores

Both models are MIT-licensed open weights, and both are expensive to run. The shapes differ.

• flash, FP8 — 578 GB of weights, served at tensor parallelism 8 with expert parallelism 8 on eight H200s, or TP4/EP4 on four GB300s.

• flash, MXFP4 — 291.72 GiB, round-to-nearest quantisation with no calibration, recommended on four H20-3e GPUs at TP4/EP4, with expert activation at W4A8 or W4A16.

• air-VL — 249.7 GB of bfloat16, tensor parallelism 4 on four 141 GB-class GPUs, or TP8 on eight 80 GB H100/H800s. No quantised build exists yet.

So the vision model is not the cheap one in absolute terms — 250 GB is still a four-GPU deployment. It is cheap relative to flash, whose FP8 build needs twice the hardware. But flash has an official quantisation that brings it to 292 GiB, and air-VL has nothing smaller than bfloat16 four days after upload. Until someone publishes a GGUF or an FP8 build for air-VL, the honest hardware comparison is 292 GiB against 250 GB, not 578 against 250.

This is the point where routing economics become concrete. Neither LingSante model is in the OrcaRouter catalogue today, so there is no price to compare. When a hosted price does appear — flash already links a third-party hosted endpoint from its card, air-VL links nothing — the structure that matters is one key across both tiers, with provider list price passed through at 0% markup and failover if a provider degrades. A pipeline that drafts with the vision model and reasons with the text model is two calls, and two calls should not mean two vendors, two contracts and two retry policies.

Where each one belongs in a workflow

Strip away the scorecards and the decision is about where the image enters.

• Intake and triage, image in, structured text out — LingSante-3.1-air-VL. It is the only one of the two that accepts an image, and its document and radiograph rows are the strongest part of its card.

• Case reasoning over text, guideline lookup, structured extraction from records — LingSante-3.1-flash. Six shared rows, six wins, and the largest gap is on the test closest to real clinical judgment.

• Long-context work over records — air-VL today, because its long-context recipe is published, despite the smaller advertised number.

• Cheapest path to a working demo — flash's MXFP4 build on four H20s is the only quantised artefact in the family, and it is three days newer than the parent release.

• Anything requiring both — two models, two stages, one pipeline. There is no row on either card that shows a single model doing perception and judgment at frontier level.

What neither card settles

Three things are unresolved across the whole family, and they are the same three for both models.

First, there is no independent evaluation. Not one row of either table has been reproduced outside Ant Group. Flash's predecessor has thousands of downloads and still no third-party score of its own; air-VL has zero downloads and a couple of likes.

Second, neither is a medical device, and both cards say so in almost identical language. That is not boilerplate to skip. It means the burden of validation sits with whoever deploys, and the benchmark tables are not a substitute for a held-out clinical set.

Third, the composite claims are soft. flash's card says the model ranks "second overall behind only GPT-6 Astra" on medical professional behaviour, without naming the aggregate behind that sentence. air-VL makes no equivalent claim, which is a point in its favour as a document even if it says less.

There is also a family-timing question. Ant's newest general flagship, Ling-3.2-flash, shipped on 5 November as a hosted, natively multimodal model with no open weights. The Sante line is still built on 3.1 and 3.0 bases. A medical model on the 3.2 backbone is the obvious next release, and its absence is the reason not to over-invest in either of these two yet.

If you need a medical model that reads images this month, LingSante-3.1-air-VL is the only open-weight option in this family and the recipe is real. If you need one that reasons, LingSante-3.1-flash wins every row you can actually compare — by sixteen points where it counts most. What no reader can do today is verify either claim, and that is the part of this comparison that will change first.

การเปรียบเทียบในบทความนี้2

ตรวจพบจากบทความนี้ · เบนช์มาร์ก: Artificial Analysis · อัปเดตทุกวัน