AntAngelMed, jointly developed by the Health Information Center of Zhejiang Province, Ant Healthcare, and Zhejiang Anzhen'er Medical Artificial Intelligence Technology Co., Ltd.(浙江省安诊儿医学人工智能科技有限公司), is the largest and most powerful open-source medical language model to date.
Core Highlights
🏆World-leading Performance on Authoritative Benchmarks: AntAngelMed surpasses all open-source models and a range of top proprietary models on OpenAI's HealthBench, and ranks first overall on the Chinese authority benchmark MedAIBench.
🧠Advanced Medical Capabilities: AntAngelMed achieves its professional medical capabilities through a rigorous three-stage training pipeline: continual pre-training on medical corpora, supervised fine-tuning with high-quality instructions, and GRPO-based reinforcement learning. This process equips the model with deep medical knowledge, sophisticated diagnostic reasoning, and robust adherence to safety and ethics.
⚡Extremely Efficient Inference: Leveraging Ling-flash-2.0’s high-efficiency MoE, AntAngelMed matches the performance of ~40B dense models while activating only 6.1B parameters of its 100B parameters. It achieves over 200 tokens/s on H20 hardware and supports 128K context length.
📊 Benchmark Results
HealthBench
HealthBench is an open-source medical evaluation benchmark released by OpenAI, designed to assess the performance of Large Language Models (LLMs) in real-world medical environments through highly simulated multi-turn dialogues. AntAngelMed achieved outstanding performance on this benchmark, ranking first among all open-source models, with a particularly significant advantage on the challenging HealthBench-Hard subset.
Figure | AntAngelMed ranks first among open-source models on HealthBench
MedAIBench
MedAIBench is an authoritative medical LLM evaluation system developed by the National Artificial Intelligence Medical Industry Pilot Facility. AntAngelMed also ranks among the top level and demonstrates strong comprehensive professionalism and safety, especially in medical knowledge Q&A and medical ethics/safety.
Figure | AntAngelMed ranks among the top level on MedAIBench
MedBench
MedBench is a scientific and rigorous benchmark designed to evaluate LLMs in the Chinese healthcare domain. It comprises 36 independently curated evaluation datasets and covers approximately 700,000 samples. AntAngelMed ranks first on the MedBench leaderboard and leads across five core dimensions: medical knowledge question answering, medical language understanding, medical language generation, complex medical reasoning, and safety and ethics, highlighting the model's professionalism, safety, and clinical applicability.
Figure | AntAngelMed ranks first on the MedBench leaderboard.
🔧 Technical Features
Professional three-stage training pipeline
AntAngelMed employs a carefully designed three-stage training process to deeply integrate general capabilities with medical expertise:
Continual Pre-Training: Based on Ling-flash-2.0, AntAngelMed is continually pre-trained with large-scale, high-quality medical corpora (encyclopedias, web text, academic publications), injecting profound domain and world knowledge.
Supervised Fine-Tuning (SFT): A multi-source and heterogeneous high-quality instruction dataset is constructed at this stage. General data (math, programming, logic) strengthen core chain-of-thought capabilities of AngAngelMed, while medical scenarios (doctor–patient Q&A, diagnostic reasoning, safety/ethics) provide deep adaptation for improved clinical performance.
Reinforcement Learning (RL): Using the GRPO algorithm and task-specific reward models, RL precisely shapes model behavior—emphasizing empathy, structural clarity, and safety boundaries, and encouraging evidence-based reasoning on complex cases to reduce hallucinations and improve accuracy.
Figure | Professional three-stage training pipeline
Efficient MoE architecture with high-speed inference
AntAngelMed inherits Ling-flash-2.0’s advanced design. Guided by Ling Scaling Laws, the model uses a 1/32 activation-ratio MoE and is comprehensively optimized across core components, including expert granularity, shared expert ratio, attention balance, no auxiliary loss + sigmoid routing, MTP layer, QK-Norm, and Partial-RoPE.
These refinements enable small-activation MoE models to deliver up to 7× efficiency over similarly sized dense architectures. In other words, with only 6.1B activated parameters, AntAngelMed can match ~40B dense model performance. Because of its small activated parameter count, AntAngelMed offers substantial speed advantages:
On H20 hardware, inference exceeds 200 tokens/s—about 3× faster than a 36B dense model.
With YaRN extrapolation, it supports a 128K context length; as output length grows, relative speedups can reach 7× or more.
Figure | Model Architecture Diagram (https://huggingface.co/inclusionAI/Ling-flash-2.0)
We have also specifically optimized AntAngelMed for inference acceleration by employing FP8 quantization combined with EAGLE3 optimization. Under a concurrency of 32, this approach significantly boosts inference throughput compared to using FP8 alone, with improvements of 71% on HumanEval, 45% on GSM8K, and as high as 94% on Math-500. This achieves a robust balance between inference performance and model stability.
1curl -s http://localhost:${PORT}/v1/chat/completions \
2 -H "Content-Type: application/json" \
3 -d '{"model": "auto", "messages": [{"role": "user", "content": "What should I do if I have a headache?"}]}'
1@misc{AntAngelMed,
2 title={AntAngelMed: A High-Performance Medical Language Model with Efficient MoE-Powered Clinical Reasoning},
3 author={AntAngelMed Team},
4 year={2025},
5 url={https://huggingface.co/MedAIBase/AntAngelMed},
6}
