Ant Group open-sources LingSante-3.1-air-VL, a 124B MoE medical multimodal model
jacek2023 · reddit · 2026-11-23
Ant Group's healthcare team open-sourced LingSante-3.1-air-VL on Hugging Face, its first large open medical multimodal model: a 124B-parameter MoE (5.5B active per token) built on Ling-3.0-flash-VL, with text+image input and a 256K context window.
- Medical imaging: identifies 8–9 positive finding types on a single complex chest X-ray (cardiomegaly, effusion, pneumonia, pneumothorax, etc.) and writes Findings/Impression sections
- Health assessment: covers 120+ skin diseases with multi-turn history taking, dietary/nutrition analysis, and lab report interpretation
- Medical search: image-to-image retrieval of papers, guidelines and similar cases, with traceable evidence chains
- Medical GUI agents: booking appointments and purchasing medication in real apps with changing UIs
Joint training with general multimodal replay preserves OCR, document/chart understanding and tool use. The team stresses it's a research model, not a medical device.
More from Models
- 55-Page Riemann Hypothesis Proof Guide, De-Formalized from Lean with Gemini and GPT — chaumian · 2026-11-23
- Sarvam open-sources its latest 105B model; founder says scaling agentic RL 'is not for the faint hearted' — itsOmSarraf_ · 2026-11-23
- Gemini 4 Neon xhigh costs 8.6x more but adds just 1-4 points on most benchmarks — jeremyphoward · 2026-11-23
- Poland's 9B-parameter Thalamus Reasoning tops LongevityBench, beating 29 models — Dr_Singularity · 2026-11-23
- Dev: Gemini 4 Neon rewrote my training kernels and supervised runs; rumors point to bigger models — xennygrimmato_ · 2026-11-23
- Inkling 2 Mini: a 26B local model that runs on laptops, and someone rebuilt Apple Notes around it — simonguozirui · 2026-11-23