A generated title card reading 'LingSante-3.1-air-VL' with the subtitle 'A 124B medical vision MoE, uploaded with no announcement'. Three icons label cards reading '249.7 GB of weights', '26-row vendor scorecard' and '0 downloads'. A footer reads 'Repository created 19 November 2026; no vendor post since 5 November.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Engineering & Research

LingSante-3.1-air-VL: Medical Vision, Shipped Without a Word

작성자

jinhao.song

게시일

최신 모델 · 20모든 모델 보기 →
벤치마크: Artificial Analysis · 매일 업데이트
모든 게시물로 돌아가기

Type LingSante-3.1-air-VL into a search engine on 23 November 2026 and the results are about something else. There is no vendor blog post, no press release, no benchmark write-up, no Reddit thread, no Chinese-language aggregator item. The most relevant thing the index can find is a different Ling model from a different week. The model itself, however, is not a rumour: huggingface.co/inclusionAI/LingSante-3.1-air-VL holds 249.7 GB of bfloat16 weights across fifty safetensors shards, a configuration file, a vision tower, custom modelling code, and an MIT licence. The repository was created at 04:31 UTC on 19 November 2026, one minute after its text-only sibling. At the time of writing it has two likes and zero downloads.

That combination — a complete, servable, open-weight release with no announcement whatsoever — is the first thing a reader needs to know, because it changes what the article can honestly claim. Nothing here has been independently evaluated. Every performance number below comes from the vendor's own benchmark image, and the vendor has not yet said a word about the model in public. What the repository does support is a sharper observation than "new model drops": on its own twenty-six-row scorecard, LingSante-3.1-air-VL finishes first or second on thirteen rows, and all thirteen involve an image. On the six text-only medical tests it never finishes better than fourth. This is a model built to look, not to reason.

What the repository proves

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' with F32 and BF16 tensor types, the notice 'This model isn't deployed by any Inference Provider', and the LingSante collection listed with 6 items updated about 13 hours ago; introduction stating 124B parameters, 5.5B active parameters per token, a 256K-token context window and the claim to be the first large multimodal medical model Ant Group's healthcare team has open-sourced; a research-only, not-a-medical-device notice; and the top of the embedded benchmark chart.

Start with what is checkable by opening files rather than trusting prose.

• Size — 124,848,460,496 parameters in the index file, 124.0 billion in bfloat16 and 164,960 in float32. The card rounds to 124B, and the Hub sidebar rounds further to 125B. Unlike the 560B sibling, where the sidebar and the card disagree by 12B because of a speculative-decoding head, these two figures agree.

• Shape — a mixture-of-experts decoder with 42 layers, the first two dense and the rest routed. 512 routed experts, eight activated per token, hidden size 2,560, vocabulary 157,184. Attention is multi-head latent attention with a 512-rank KV compression, and there is no separate multi-token-prediction block in the counted parameters.

• Vision path — a 27-layer ViT with hidden size 1,152, patch size 16, 2×2 spatial merging, and a 4,096-dimension projector into the language model. The config calls the tower qwen3_moe_vit, which is a component name rather than a statement about whose model this is.

• Tokens — the tokenizer reserves an image token, a video token, a vision-start and a vision-end token, and the config carries a three-way multimodal rotary split of [8, 12, 12]. The card advertises text and images; video plumbing exists in the configuration but is not claimed anywhere in the prose.

• Lineage — the card states the model is built on Ling-3.0-flash-VL, inclusionAI's September vision model. That matters for expectations: this is the medical specialist derived from a mid-sized general vision model, not a trimmed-down version of the 560B flagship.

• Licence — MIT, with no acceptable-use rider. The card adds a plain-language restriction of its own: a research model, not a medical device, not a basis for diagnosis or treatment.

• Files — weights, config, tokenizer, chat template, image and video processing code, two modelling modules, and the README. No training-data manifest, no evaluation harness, no technical report, no paper.

124B total, 5.5B active: the air tier

The "air" in the name is doing real work. The card claims 124B total parameters and 5.5B active per token, which puts it in the same bracket as the September Ling-3.0-flash-Sante release — 124B total, 5.1B active — but with a vision encoder attached and a newer base. It is the light tier of a medical family whose heavy tier is LingSante-3.1-flash at 560B total and 25B active.

Active parameters are the ones that determine compute per token, and they are not the ones that determine the memory bill. At 5.5B active this is a model that decodes at roughly the speed of a mid-sized dense model, while occupying 250 GB of accelerator memory. That is the trade the architecture makes: cheap arithmetic, expensive residency. For a hospital group that already owns four large GPUs, it is a good trade. For anyone renting by the hour, the arithmetic is less friendly, and the absence of a published hosted price makes it impossible to say whether buying tokens would be cheaper.

Four GPUs, 250 GB, one serving recipe

Quiet uploads usually ship weights and nothing else. This one ships an operational recipe, which is the strongest available signal that somebody intends to run it.

• SGLang — the recommended path is a day-zero image, lmsysorg/sglang:dev-Ling-3.0-flash-VL, with tensor parallelism 4 on four 141 GB-class GPUs. On eight 80 GB H100s or H800s, switch to tensor parallelism 8. The engine supports NEXTN speculative decoding and the usual Hopper decode flags.

• Context — the launch command sets --context-length 262144 and passes a YaRN block with factor 2.0 over an original maximum of 131,072, with the same 6,000,000 rope theta and 0.5 partial rotary factor the config declares.

• vLLM — a second path through inclusionAI's own fork, vllm-ling-v3, which requires --trust-remote-code because the modelling code lives in the repository rather than in upstream transformers. The recipe enables prefix caching and an align mamba cache mode.

• Parsers — both paths register a ling3 reasoning parser and a ling3 tool-call parser, so the model's thinking traces and tool calls are separable from its answer.

This is where the release stops being a file dump. A team that did not intend to serve the model would not have written the YaRN invocation, checked it against a four-GPU topology, and wired the tool parser.

Its own scorecard is an eye test

A generated two-panel infographic titled 'LingSante-3.1-air-VL - where it wins'. The left panel, headed 'Image rows', shows the figure '13' with 'first or second place' and the line '8 firsts, 5 seconds, out of 20 image benchmarks'. The right panel, headed 'Text-only rows', shows '4th' with 'best finish of six rows' and the line 'HealthBench Professional 49.59'. A footer reads 'Vendor-reported figures from the 26-row model card, 23 November 2026; no independent run.' The OrcaRouter logo is composited in a white strip along the bottom edge.

The evaluation section is a single image: twenty-six benchmark rows across eight columns, with the comparator set drawn from cheap open models — GLM-5.3-Flash, DeepSeek-V4.1-Flash-1017, MiniMax-M3.1-Flash — plus three frontier systems at their highest reasoning setting, GPT-6.1 Sol, Claude Opus 5.5 and Gemini 3.8 Flash, and Kimi-K3.1-Pro at 2.8 trillion parameters. The card says all comparators run at the highest available effort, and LingSante-3.1-air-VL is evaluated with thinking on at temperature 0.6, top-p 0.95, top-k 20. Vendor-run throughout.

The wins are concentrated, and they are visual. On chest radiographs it takes CheXpert Plus at 46.21 against GPT-6.1 Sol's 44.76, MIMIC-CXR at 47.20 against 45.21, and IU X-RAY at 64.34 against Kimi-K3.1-Pro's 56.88. On medical visual question answering it takes OmniMedVQA at 93.64 against 86.53, and VQA-RAD at 81.15 against Claude Opus 5.5's 78.94. On document parsing it takes OmniDocBench v1.5 at 91.20, ahead of Kimi at 91.05 and Claude at 90.56. It is second on three more image rows — PathVQA, PMC-VQA, SLAKE — each time behind Claude Opus 5.5, not behind a small open model.

Three of the first-place rows are in-house evaluations the field cannot reproduce: AFUMED-Derm at 70.01, AFUMED-Report at 90.70, and AFUMED-Food at 72.80, where Gemini 3.8 Flash edges it at 73.13. Those datasets belong to Ant's healthcare team. They are useful evidence that the model was trained toward skin conditions, diet and report interpretation, and they are not evidence that it beats anyone on a test anyone else can run.

Where the table turns against it

The text-only block reads like a different model. MedQA-USMLE: 94.34, last of eight. MedMCQA: 81.54, seventh. HealthBench Professional: 49.59, sixth, against 65.72 for its own 560B sibling, 62.79 for GPT-6.1 Sol and 61.19 for Claude Opus 5.5. MedXpertQA-Text: 50.65, fifth. ReDisQA: 90.81, sixth. DiagnosisArena-MCQ: 81.70, fourth — its best text result and still behind three comparators.

The pattern also shows up on tests that mix perception with knowledge. MMMU-Pro-Med is 73.95, fifth of eight. GMAI-MMBench is 66.33, fourth, behind Gemini at 72.55 and Claude at 71.60. CharXiv-RQ, which asks a model to reason over scientific charts, is 84.30 against Claude's 92.90. The model can read the pixels; the step from pixels to clinical judgment is where the gap opens.

That is not a defect so much as a design statement. A 5.5B-active budget is enough to encode what a radiograph shows and not enough to carry the world knowledge that MedQA rewards. Anyone planning to replace a reasoning model with this one is reading the wrong rows.

256K context, and this time it is documented

Context claims are where quietly shipped models usually go soft. LingSante-3.1-flash is the cautionary example: a one-million-token promise in the introduction, a 262,144-token configured window in the shipped config, and no published recipe explaining the gap.

LingSante-3.1-air-VL does the opposite. The config's max_position_embeddings is 131,072, and the quickstart explicitly doubles it to 262,144 with a YaRN factor of 2.0 and the flag that permits overwriting a longer context length. So 256K is not a benchmark-harness capability here; it is a server argument you can copy. The practical caveats remain: long-context support is not the same as complete retrieval, and the card says so itself.

What zero downloads means

Two likes, zero downloads, four days old. Those numbers are consistent with a model that nobody has noticed, and also consistent with a model that a small number of people have pulled through a mirror that does not count. They are not evidence of quality in either direction. The counter did move from one to two while this piece was being written, which is a reminder that it measures attention rather than use.

The repository is also still being worked on. Its twelfth commit landed at 17:03 UTC on 22 November, from a contributor account rather than an organisation bot, and its message describes adding local benchmark assets and updating the model card. Somebody is polishing the listing. The model was added to inclusionAI's LingSante collection the same day, alongside Ling-3.0-flash-Sante and LingSante-3.1-flash. None of that is an announcement, but it is not the behaviour of an accidental upload either.

What is missing is more concrete than a download counter. There is no hosted endpoint. The card links no API, unlike its flash sibling, which points to a third-party hosted service. There is no price, no rate limit, no context-tier table. There is no entry on a neutral leaderboard and no independent run of a single row. And there is no announcement: the vendor's English blog index still ends at 5 November, with Ling-3.2-flash, and the medical releases of 19 November appear in neither the blog nor the changelog.

Screenshot of the Ant Ling blog index at ant-ling.com/en/blog captured 23 November 2026 in English: a banner reading 'Ling-3.2-flash Is Now Live', then Model Family cards for Ling (Language) and Ming (Multimodal), then a 'What's New' section listing four announcements - Ling-3.2-flash: Native Multimodal Text Model Now Live dated 2026-11-05, Ling-3.1-flash: A Model Upgrade for Real-World Tasks dated 2026-09-30, Ming-Image-0.1-Design Series Now Open Source dated 2026-09-23, and Ling-3.0-flash-VL Open Source Release dated 2026-09-09. Nothing dated between 6 and 23 November appears.

There is also a family question worth stating plainly. Ant's newest flagship, Ling-3.2-flash, is natively multimodal and closed — hosted, not open-weight. LingSante-3.1-air-VL is the open-weight branch, and it sits on the 3.0-generation vision base rather than the 3.2 one. Open weights and current generation are, for now, two different models.

When a quiet upload becomes a release

Ant has done this before. Ling-3.1-flash appeared as weights first and a developer blog later; LingSante-3.1-flash landed the same way one minute before this model. A missing announcement is therefore weak evidence that none is coming, and the honest description today is a checkpoint rather than a product.

The pipeline this model implies is a two-model pipeline, and that is the useful thing to take away. A medical vision model produces findings, measurements, structured text. Something else has to weigh those findings against guidelines and produce a judgment. That is precisely the shape a routing layer exists for: OrcaRouter's routing DSL composes several models into a single call, so the vision stage and the reasoning stage can come from different vendors behind one key, with automatic failover if a provider degrades. LingSante-3.1-air-VL is not in the OrcaRouter catalogue today, and this article does not pretend otherwise; the point is that even at its best, the card describes half a system.

What to watch is short and specific. A post on the Ant Ling developer blog. A model entry on a neutral leaderboard. An independent run of any row, especially HealthBench Professional. A hosted price. A community quantisation that gets the footprint under a single 80 GB card. Any one of those turns this from a describable checkpoint into a testable model, and the first one to land will also tell you whether the vendor considers it finished.

Until then, the fair summary is narrow and useful. LingSante-3.1-air-VL is a real, complete, MIT-licensed 124B medical vision mixture-of-experts with a documented 256K serving recipe and a vendor scorecard on which every strong result has an image attached. It looks like the perception tier of a medical model family, published before anyone was ready to talk about it.

이 글에서 비교한 모델1

이 글에서 자동 인식 · 벤치마크: Artificial Analysis · 매일 업데이트