Ant Group has released an open medical model that it says beats the top closed models on several health benchmarks. Its Ant Ling team announced LingSante-3.1-flash on X at 15:22 UTC on November 22, 2026: “an open-weight MoE model with 560B total parameters and 25B activations per token.” The team’s framing: “LingSante-3.1-flash moves beyond medical QA toward real-world healthcare problem.”
The weights are on Hugging Face under an MIT license. Here is what it is, where Ant’s own charts show it winning and losing, and what its model card warns.
On this page · 6 sectionsOpen
Ant Group’s Ant Ling team released LingSante-3.1-flash on November 22, 2026: an open-weight mixture-of-experts model with 560B total and 25B active parameters.
The weights are on Hugging Face under an MIT license, with a 1M-token context window, aimed at medical reasoning and long clinical workflows.
On Ant’s own chart it leads Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon on HealthBench Professional, HLE with tools and BrowseComp, and edges GPT-6 Astra on EHR-Complex.
The same chart shows it trailing on MedXpertQA, MedEthicAlign, MCP-Atlas and PhysicianBench, and every score is the lab’s own.
A second model, LingSante-3.1-air-VL (124B total, 5.5B active), reads medical images; the model card says neither should diagnose or prescribe on its own.
§ 01What Ant released
| Item | Value |
|---|---|
| Architecture | Mixture of experts |
| Total parameters | 560B |
| Active per token | 25B |
| Context window | 1M tokens |
| License | MIT |
| Repository created | November 19, 2026 |
Two hours later Ant added LingSante-3.1-air-VL, a multimodal model with 124B total and 5.5B active parameters for medical images, report writing and medical reasoning. The announcement had 246,868 views when we read it at 20:02 UTC.
§ 02Where it leads, and where it trails
Ant compares the model with DeepSeek, GLM, Kimi, Claude Opus 5.5, GPT-6 Astra and Gemini 4 Argon. Against the three US frontier models, the scores read off its chart split both ways:
| Benchmark | LingSante-3.1-flash | Best of Claude, GPT-6 Astra, Gemini | Result |
|---|---|---|---|
| HealthBench Professional | 65.72 | 64.18 (GPT-6 Astra) | Leads |
| HLE, text only, with tools | 72.22 | 63.59 (GPT-6 Astra) | Leads |
| BrowseComp, with context | 94.10 | 91.50 (GPT-6 Astra) | Leads |
| EHR-Complex | 74.40 | 74.20 (GPT-6 Astra) | Leads by 0.2 |
| DiagnosisArena-MCQ | 83.9 | 84.7 (GPT-6 Astra) | Trails |
| PhysicianBench | 57 | 58 (Gemini 4 Argon) | Trails |
| MedXpertQA-Text | 57.27 | 71.35 (Gemini 4 Argon) | Trails |
| MedEthicAlign | 80.85 | 88.08 (Claude Opus 5.5) | Trails |
| MCP-Atlas | 83.98 | 88.30 (Gemini 4 Argon) | Trails |
Ant’s own X post names PhysicianBench (57) as a strong result; on the same chart Gemini 4 Argon scores 58. The headline claim holds for the three benchmarks Ant names, HealthBench Professional, HLE and BrowseComp, and the margin on EHR-Complex is 0.2 points.
§ 03What the model card warns
Ant’s card is unusually direct about limits. “The model does not replace qualified healthcare professionals. It must not independently diagnose, prescribe, determine treatment, or make clinical triage decisions.” It adds: “Benchmark and simulated-environment results do not establish clinical effectiveness.” And on the 1M-token window: “Long-context support does not guarantee complete retrieval of every relevant fact.”
The card lists the intended uses as medical question answering and clinical document analysis with qualified human review.
§ 04What is missing
- Independent scores. Every number above is Ant’s. Several benchmarks on its chart (AFUMED, MedEthicAlign, InclusionMed) are in a section Ant labels internal.
- Running costs. The card gives serving setups for large GPU clusters, not a price or a hosted API.
- Clinical evidence. The card itself says benchmark results do not establish clinical effectiveness.
§ 05What we are watching
- Outside evaluations. HealthBench runs by groups other than Ant.
- Hosted access. Whether OpenRouter or other providers list it, and at what price.
- Rivals’ replies. Open medical models from other labs, and any update from Google, OpenAI or Anthropic on health benchmarks.
§ 06Sources
- Ant Ling on X, 15:22 UTC, November 22, 2026, with its benchmark chart
- Ant Ling on LingSante-3.1-air-VL, 17:26 UTC, November 22, 2026
- LingSante-3.1-flash model card on Hugging Face, read November 22, 2026
Q1When was it released?
Ant Ling announced it on X at 15:22 UTC on November 22, 2026. The Hugging Face repository was created on November 19 and last updated on November 22.
Q2What license is it under?
MIT, according to its Hugging Face model card.
Q3What is LingSante-3.1-air-VL?
A smaller multimodal model, 124B total and 5.5B active parameters, for medical images, report writing and medical reasoning, announced at 17:26 UTC the same day.
Q4Are the benchmark scores independent?
No. All scores in this post come from Ant’s own charts. Several benchmarks, such as AFUMED and InclusionMed, are labelled by Ant as internal.
Q5Does CellCog use LingSante?
No. CellCog does not route to LingSante, and CellCog is not offered for patient health information.
