Ant’s 560B Medical AI Leads 4 of 16 Tests in Its Own Chart

Ant Group released LingSante-3.1-flash on Sunday, an open-weight medical AI model with 560 billion parameters, of which 25 billion are active per token. The weights are on Hugging Face under an MIT license.

In Ant’s own benchmark chart, the model has the top score on 4 of 16 tests against six rivals, including GPT-6 Astra, Claude Opus 5.5 and Gemini 4 Argon, by our count. It trails the leader on the other 12.

Ant’s launch post said the model leads on HealthBench Professional, HLE and BrowseComp. These are Ant’s own results, not independent tests. Our AI model benchmarks tracker follows how the leaders compare.

Business Pill · The Model and the Harness

An evergreen Business Engineer explainer. It teaches the general idea only and does not discuss this release.

The key insight: Ant Group’s open-weight LingSante-3.1-flash uses 25B of its 560B parameters per token (4.5%, our arithmetic) and has the top score on 4 of 16 tests in Ant’s own chart, leading GPT-6 Astra on HealthBench Professional by 1.54 points while trailing Gemini 4 Argon by 14.08 on MedXpertQA-Text; the results are Ant’s own.

What Ant Released

The model is built on Ant’s general Ling-3.1-flash and keeps its design: a mixture-of-experts model with a 1-million-token context window, per the model card.

Only 25B of its 560B parameters work on each token, about 4.5%, by our arithmetic.

Its predecessor, Ling-3.0-flash-Sante, had 124B parameters with 5.1B active. The new model is 4.5 times larger in total and 4.9 times in active parameters, by our arithmetic.

LingSante-3.1-flash leads the next-best rival by 0.20 to 2.22 points on the four tests where it tops Ant&#8217
LingSante-3.1-flash leads the next-best rival by 0.20 to 2.22 points on the four tests where it tops Ant’s chart, by our arithmetic on Ant’s figures.

Where It Leads, in Ant’s Chart

On HealthBench Professional it scores 65.72 against 64.18 for GPT-6 Astra, a 1.54-point lead.

On HLE, text-only with tools, it scores 72.22 against 70.00 for Kimi-K3.1-Pro. On BrowseComp, with context, it scores 94.10 against 92.10.

On EHR-Complex, a test Ant groups with long-horizon medical workflows, it scores 74.40 against 74.20 for GPT-6 Astra, a 0.20-point lead.

Where It Trails

On MedXpertQA-Text it scores 57.27, 14.08 points behind Gemini 4 Argon’s 71.35. On Ant’s own AFUMED-SearchTrace test it trails GPT-6 Astra by 0.06 points.

Bar chart: LingSante-3.1-flash has the top score on 4 of 16 benchmarks in Ant's chart and trails the leader on 12. Data: Ant Group
LingSante-3.1-flash has the top score on 4 of the 16 tests in Ant’s chart and trails the leader on 12, by our count.

It scores above DeepSeek-V4.1-Flash, a model of similar total size in the chart, on all 16 tests. Four of the 16 are Ant’s internal benchmarks.

The Structural Read

As we read it, Ant’s pitch is breadth at low running cost: an open model with 4.5% of parameters active that scores near the top on medical tasks, by its own chart.

As we read it, the leads are narrow, from 0.20 to 2.22 points, so independent tests will matter more than the launch chart.

Ant Group, LingSante-3.1-flash model card

“It is not a medical device and must not be used as an independent basis for diagnosis, prescriptions, or treatment.”

Three Implications

FOR HOSPITALS AND HEALTH-TECH BUILDERS The model card says the model is for research and evaluation and is not a medical device.

FOR AI LABS Ant’s chart shows narrow leads on 4 of 16 tests and gaps of up to 14.08 points elsewhere, by our arithmetic.

FOR INVESTORS The weights are free under an MIT license, with 25B of 560B parameters active per token.

What Is Not Established

The scores come from Ant. Its model card and launch post cite no outside tests.

Ant’s model card says the model “is not a medical device and must not be used as an independent basis for diagnosis, prescriptions, or treatment.” It adds that benchmark results “do not establish clinical effectiveness.”

A self-published report by researcher Will DePue said the base model, Ling-3.1-flash, carries a fingerprint matching Nvidia’s Blackwell chips, as we reported on 15 Nov. The report says it does not establish how access was obtained.

This Week In The AI Supercycle

A World Without an Order, and an AI Central Bank Without a Successor

This week’s issue: the week AI memory stayed largely sold out while the people holding the shares paid: why the AI Supercycle is not a dot-com replay, read through geopolitics, macro and technology, with the gap between building and financing the widest we’ve recorded.

Read the full issue →

Business Engineer Framework

The AI Supercycle

The Business Engineer map of who builds, finances and captures value in AI.

Read the Map of AI →

The Bottom Line

Ant’s open-weight medical model tops 4 of 16 tests in its own chart, by our count, while using 25B of its 560B parameters per token. The results are Ant’s own. What to watch is whether hospitals or independent evaluators publish results that match Ant’s.

97,000+ executives read Business Engineer for the AI strategy frameworks cited by ChatGPT, Claude, and Perplexity.

A note on sourcing. Model details, quotes and every benchmark score are from Ant Group’s Hugging Face model card and its benchmark chart, and from Ant’s launch post on X. Scores were read from Ant’s chart as printed; they are Ant’s own results, not independent tests. The count of 4 of 16, the 4.5%, the 4.5x and 4.9x, and the point gaps are our arithmetic. The Blackwell fingerprint claim is from Will DePue’s self-published report, as we reported on 15 Nov.

Sources: Hugging Face model card, inclusionAI/LingSante-3.1-flash (Ant Group), incl. benchmark overview chart, read 22 Nov 2026 · Ant Group (@AntLingAGI), launch post on X, 22 Nov 2026 · Rohan Paul, post on X, 22 Nov 2026 · FourWeekMBA, DePue Blackwell fingerprint report, 15 Nov 2026

Scroll to Top
FourWeekMBA