A generated title card reading 'LingSante-3.1-flash' with the subtitle '560B of open medical weights, one outside score'. Three rounded cards read 'HealthBench Professional 65.72, vendor-reported', 'MLCR-AA 23.3%, independent' and 'Announced 22 November 2026'. A footer reads 'Vendor figures from Ant's 41-row card; MLCR-AA per Artificial Analysis.' The OrcaRouter logo is composited in a white strip along the bottom edge.
Engineering & Research

LingSante-3.1-flash Tops HealthBench — the Only Outside Run Says 23.3%

Autore

Gideon Frost

Data di pubblicazione

Ultimi modelli · 20Vedi tutti i modelli →
Benchmark: Artificial Analysis · aggiornato ogni giorno
Torna a tutti gli articoli

At 15:22 UTC on 22 November 2026, the Ant Ling account posted three messages announcing LingSante-3.1-flash, "an open-weight MoE model with 560B total parameters and 25B activations per token" that "unifies diverse medical capabilities while retaining general capabilities." The model card's own comparison table puts it ahead of GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon on HealthBench Professional, HLE with tools and BrowseComp with context. The weights had been sitting on Hugging Face since 04:30 UTC on 19 November, unannounced, with one like on the repository.

By the morning of 23 November the like count was 27 and the download count was still zero. No developer-blog post existed. And the one score on the model that no vendor produced — a medical long-context reasoning benchmark run by Artificial Analysis — read 23.3%.

That is the shape of this release: a genuinely large, genuinely open medical model, described by its maker with numbers that are the best in their rows, and almost entirely unexamined from outside. The interesting question is not whether open models have caught closed ones in medicine. It is which parts of the claim survive contact with the table.

What Ant announced on 22 November

The announcement thread is unusually specific for a launch post, which makes it easy to check. The claims:

• HealthBench Professional — 65.72, described as the best score in the comparison.

• HLE, text only, with tools — 72.22 for a single agent, 74.87 after multi-agent aggregation.

• BrowseComp with context — 94.10.

• PhysicianBench — 57, and PatientAgentBench — 4.722 on a five-point scale.

• "Second overall behind only GPT-6 Astra" on a three-benchmark medical professional-behaviour average.

The card supports the first two outright. On the 41-row table, 65.72 is the highest HealthBench Professional figure, ahead of GPT-6 Astra at 64.18, Claude Opus 5.5 at 61.19 and Gemini 4 Argon at 55.91. HLE with tools is 72.22 in single-agent mode against Kimi-K3.1-Pro's 70.00, Tencent Hy4's 65.76 and GPT-6 Astra's 63.59. BrowseComp is 94.10 against Kimi's 92.10 and Astra's 91.50. So three headline rows really do favour this model over much larger closed systems.

The "second overall" sentence is the softer claim. There is no published weighting for that average. The three rows nearest to it — HealthBench Professional, PhysicianBench and PatientAgentBench — rank it first, second and first respectively, so a composite is plausible. It is not reproducible from the card, and until the weights of that composite are published it should be read as a vendor framing rather than a measurement.

Dove il modello aperto vince davvero

The medical-agent rows are the strongest part of the release, and they are not decorative. On EHR-Complex, which measures multi-step work inside structured health records, LingSante-3.1-flash scores 74.40 against GPT-6 Astra's 74.20 and Claude Opus 5.5's 41.80. On PatientAgentBench it scores 4.722 against Astra's 4.438. On CHI-Bench, a consumer-health interaction benchmark, it scores 28.00 against every comparator's low-twenties or worse. On MentalHealthBench it trails only GPT-6 Astra, 57.50 to 54.64.

These are the benchmarks where a 560B open-weight model beating a frontier closed system matters commercially. A hospital group that wants patient records to stay inside its own perimeter cannot send them to a closed API. If the open model is within a point of the closed one on the task you care about, the deployment argument stops being ideological and becomes arithmetic.

Screenshot of the Hugging Face model card for inclusionAI/LingSante-3.1-flash captured 23 November 2026 in English: 27 likes and a blank downloads-last-month figure; tags Text Generation, Safetensors, bailing_hybrid, conversational, custom_code and fp8 with License mit; sidebar showing 572B parameters in BF16, F32 and F8_E4M3, the line 'This model isn't deployed by any Inference Provider', and a six-item LingSante collection; introduction stating 560 billion total parameters, 25 billion per token and a 1-million-token context window; the research-and-evaluation-only notice; and a grid of benchmark chart thumbnails whose legend names LingSante-3.1-flash, DeepSeek-V4.1-Flash-1017, GLM-5.3, Kimi-K3.1-Pro, Claude Opus 5.5, GPT-6.2 Astra and Gemini 4 Argon.

The vendor also reports the serving story that makes that arithmetic possible: an MIT licence, a day-zero SGLang image, a vLLM recipe, and a 291.72 GiB MXFP4 checkpoint that runs on four H20-3e GPUs rather than eight H200s. Quantization work this specific is a signal that the team intends the model to be deployed, not merely cited.

Where it does not

Read the same table for the rows the announcement did not lead with, and the pattern is consistent: this model wins behaviour and agentic workflow rows, and loses knowledge-recall and safety rows.

• MedQA-USMLE — 94.50, against GPT-6 Astra 97.33, Gemini 4 Argon 97.33 and Claude Opus 5.5 97.25. Fourth of nine.

• MedXpertQA-Text — 57.27, against Argon 71.35 and Opus 67.14. Sixth of nine.

• MedMCQA — 83.89, against Argon 89.55 and Opus 89.53.

• MedEthiAlign — 80.85, fifth of nine, behind Tencent Hy4 88.83, Claude Opus 5.5 88.08 and GLM-5.3 85.84.

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

• Terminal-Bench 4.0 — 26.5, above only MiniMax-M3.1's 25.8 among nine models.

• GDPval-AA v2.1 — 1,600 Elo, below DeepSeek-V4.1-Flash at 1,746 and GLM-5.3 at 1,644.

The announcement's phrase "while retaining general capabilities" is therefore doing more work than the table supports. General capability is retained in the sense that the model still answers general questions and still runs tools; it is not retained at the level of the base model's agentic scores. The medical training lifted reasoning and medical workflow and cost terminal and coding-agent ability. That trade is visible in the numbers, and it is the most decision-relevant thing on the card.

The only outside run says 23.3%

Artificial Analysis lists LingSante-3.1-flash under the slug lingsante-3-1-flash-1119. As of 23 November the page shows no Intelligence Index — the field reads Unknown — and its pricing panel reads $0.00 per million tokens because the model has no hosted API in the catalogue. What the page does carry is MLCR-AA, a private medical long-context reasoning benchmark built with Wisedocs. LingSante-3.1-flash scores 23.3%.

That single number is the entire independent record. It does not contradict the vendor's 65.72 on HealthBench Professional, because the two tests measure different things: HealthBench Professional is a rubric-scored clinical conversation benchmark, MLCR-AA is long-context medical reasoning over private documents. But it does mean that the confident sentence "matching closed source ones" rests on a vendor harness, and the one outside harness that exists chose a harder task and returned a much lower number.

A generated single-column scorecard titled 'LingSante-3.1-flash — vendor claim vs outside run'. Six rows read 'HealthBench Professional: 65.72, vendor', 'HLE with tools: 72.22, vendor', 'BrowseComp with context: 94.10, vendor', 'MLCR-AA: 23.3%, independent', 'AA Intelligence Index: not listed' and 'Hugging Face downloads after four days: 0'. A footer reads 'Vendor rows from Ant's card, 22 November 2026; MLCR-AA per Artificial Analysis.' The OrcaRouter logo sits in a white strip at the bottom.

The honest summary of the evidence is asymmetric. The vendor published 41 rows across nine models, all on its own harness, with footnotes that specify the harnesses used. Nobody outside Ant Group has reproduced a single row. The download counter — zero, four days after upload and one day after the announcement — is consistent with a release that the announcement has not yet turned into trials.

What an open medical model is actually for

It helps to separate two audiences, because they need different things from this release.

A model team wants a reproducible benchmark. On that measure LingSante-3.1-flash is promising and unproven: the weights are complete, the licence is permissive, the serving path is documented, and the evaluation is single-source. Wait for an independent run before quoting a number in a paper.

A health system wants a deployable system that keeps data inside its perimeter. On that measure the release is already useful. 560B total with 25B active is a frontier-class backbone, the MXFP4 build brings the hardware floor to four H20-3e cards, and the research-only notice on the card is the same disclaimer every clinical model carries. The gap between "promising" and "usable" here is procurement, not capability.

One practical note for anyone building the comparison themselves: the closed baselines on Ant's table are routable. Claude Opus 5.5, GPT-6.2 Astra, Gemini 4 Argon, DeepSeek V4.1 Flash and GLM 5.3 Flash are all on OrcaRouter at provider list price with no markup, so a team can run its own 41-row probe against the same comparators without signing five contracts. LingSante-3.1-flash is not on the platform; it has to be self-hosted or reached through the vendor's launch partner, whose free tier currently requires a $20-per-month token plan.

Cosa cambierebbe questo quadro?

Three things, in order of likelihood. An independent benchmark run — MLCR-AA already exists, but a second one on a clinical task would matter more than another vendor row. A download count that moves, which would show hospitals actually testing it rather than bookmarking it. And a developer-blog post, which the Ant Ling site has not published, despite having used that site for launch notes through 2026.

Until one of those lands, the defensible statement is narrow and still interesting: an MIT-licensed 560B medical mixture-of-experts model, announced on 22 November, that beats closed frontier systems on several medical-behaviour and medical-agent benchmarks according to its own harness, trails them on medical knowledge and safety, and has exactly one independent data point — 23.3% on a harder test. That is not nothing. It is also not parity yet.

A generated two-column infographic titled 'Where LingSante-3.1-flash leads and trails'. The left column 'Leads, versus closed models' lists 'HealthBench Pro: 65.72 vs 64.18', 'HLE with tools: 72.22 vs 63.59' and 'BrowseComp: 94.10 vs 91.50' with green check icons. The right column 'Trails, versus closed models' lists 'MedQA: 94.50 vs 97.33', 'MedXpertQA: 57.27 vs 71.35' and 'Terminal-Bench 4.0: 26.5 vs 59.6' with amber warning icons. A footer reads 'Comparators are GPT-6 Astra and Gemini 4 Argon; all rows vendor-reported, unreproduced.'

Confrontati in questo articolo3

Rilevato da questo articolo · Benchmark: Artificial Analysis · aggiornato ogni giorno