Ant Group’s healthcare team has presented LingSante-3.1-air-VL, a text-and-image medical research model whose announcement claims leads over GPT-6.1 Sol and Gemini 3.8 Flash across four medical evaluation categories. Those comparisons come from the developers’ own evaluation, not independent validation. Source Source
The model card says: “LingSante-3.1-air-VL is a research model and is not a medical device.” It warns that outputs may be incomplete or inaccurate and must not replace professional medical advice, diagnosis or treatment. Benchmark wins are not a permission slip for patient care. Source
What the team says it beats
The announcement reports leads over both named competitors on AFUMED-Derm, AFUMED-Report, MIMIC-CXR and medical visual question answering, or VQA. That last category is an average across benchmarks, rather than a single test. Source
It also reports higher scores than GPT-6.1 Sol on BrowseComp-VL and Gemini 3.8 Flash on AndroidWorld, alongside what it describes as competitive performance on MMSearch. Those are separate competitor-and-task pairings, not a claim that it beats both models across every browsing and tool-use test. Source
The announcement post gives no numerical scores, winning margins or evaluation settings. The model card links benchmark charts. Source Source
What the model is built to do
Its model card describes a mixture-of-experts model with 124 billion total parameters, of which 5.5 billion are active per token. Hugging Face’s model-size listing gives a different figure: 125 billion. The page doesn’t explain the discrepancy. The card also specifies text and image inputs and a 256K-token context window. Source
The team describes four focus areas: medical imaging reports, multimodal health assessment, medical search and evidence synthesis, and medical app-interface agents. Intended uses include assisted report generation in research settings and assessments reviewed by qualified personnel. Search, tool use and app interaction require an external environment to provide tools, maintain state and validate actions; consequential actions require human approval. Source
The card calls the model “lightweight,” while its SGLang deployment example uses four 141GB-class GPUs. That is a fairly specialized definition of light. It is a documented example configuration, not a verified minimum requirement. Source
The team describes this as its first large multimodal medical model released as open source. That label alone doesn’t establish unrestricted usage rights; the license terms remain unverified. Source
The model card warns that benchmark and simulated-environment results do not establish clinical effectiveness. It says the model must not independently diagnose, prescribe, determine treatment or make clinical triage decisions. Source
Comments