Ant Open-Sources Medical Multimodal Model: 124B Parameters, 5.5B Activated, Can Read Scans and Write Reports
Ant Group's medical team has open-sourced LingSante-3.1-air-VL, with 124B total parameters and only 5.5B activated. It can generate chest X-ray reports, analyze skin conditions and diet, retrieve medical evidence, and operate medical apps. It outperforms GPT-6.1 Sol and Gemini 3.8 Flash on multiple medical benchmarks. The model is for research only and does not replace doctors' diagnoses.
In the open-source AI ecosystem, Ant Group's medical team has released another model: LingSante-3.1-air-VL. It is open-sourced simultaneously on Hugging Face and ModelScope and is positioned as a native multimodal MoE for real-world medical scenarios.
First, parameters. It has 124B total parameters and activates only 5.5B per token. Built on Ling-3.0-flash-VL, it supports text and image inputs and has a 256K context window. At this scale, capability does not shrink, but inference costs are much lower. To run it locally, the official SGLang recommended configuration is 4×141GB-class GPUs, or 8×80GB H100/H800.
The model's capabilities are concentrated in four clinical directions.
Medical Imaging Reports
Given a chest X-ray, the model can output a structured report containing Findings and Impression. In official demos, it can also identify 8-9 positive signs at once, including cardiomegaly, pleural effusion, atelectasis, consolidation, pneumonia, and pneumothorax.
Multimodal Health Assessment
With a skin photo and symptom history, the model can provide a preliminary assessment covering 120+ skin diseases. Dietary analysis can identify overlapping or mixed ingredients and estimate portions and nutrition. For lab reports, it links blood counts with inflammation and immune markers and gives next-step recommendations.
Medical Search and Evidence Synthesis
The model supports image-based search to retrieve medical papers, clinical guidelines, and similar cases. Before generating a report, it can use imaging tools and bounding boxes to verify lesion locations. The whole pipeline: image → search → evidence → answer, and every conclusion can be traced back to evidence.
Medical GUI Agent
The model can operate real medical apps. When registering online, it selects hospital, department, doctor, and time, and fills in patient information; when buying medicine, it searches for drugs, selects specifications, fills in prescriptions and addresses, and handles pop-ups and validation. The official team says evaluation used real applications, with dynamic interfaces and login states.
Training and Optimization
In training, the model's reinforcement learning objective considers not only literal match but also clinical significance, lesion location, false positives, and missed diagnoses. In other words, it cares more about which findings must not be missed.
Evaluation Results
Official evaluations show that on four medical benchmarks—AFUMED-Derm, AFUMED-Report, MIMIC-CXR, and medical VQA—LingSante-3.1-air-VL leads comparison models overall and exceeds GPT-6.1 Sol and Gemini 3.8 Flash on all four. General capabilities are not lost either: it is higher than GPT-6.1 Sol on BrowseComp-VL, higher than Gemini 3.8 Flash on AndroidWorld, and competitive on MMSearch.
How to Access
The weights are open-sourced. You can find inclusionAI/LingSante-3.1-air-VL on Hugging Face and ModelScope. You can also call it directly via Computrix's API, but the free quota is limited and the basic subscription is $20 per month.
The model page clearly states: this is a research model, not a medical device. Outputs may be incomplete or inaccurate and cannot replace a doctor's diagnosis or treatment recommendations. For real medical scenarios, professional review is required. This sentence is worth remembering for everyone working in medical AI.
发布时间: 2026-11-23 04:35