AntAngelMed
A medical-domain MoE large model developed by Ant Health together with the Zhejiang Provincial Health Information Center and Zhejiang Anzhen'er Medical AI. It has 103B total parameters and 6.1B activated parameters, supports a 128K context, and is released under the Apache-2.0 license, targeting localized deployment scenarios such as clinical triage, medical record summarization, and decision support.
Timeline (last 90 days)
AntAngelMed is a medical-domain Mixture-of-Experts (MoE) model with 103B total parameters and 6.1B activated parameters.
AntAngelMed was jointly developed by Ant Health together with the Zhejiang Provincial Health Information Center and Zhejiang Anzhener Medical AI.
AntAngelMed uses the Apache-2.0 license, with open weights, allowing commercial use, modification and self-hosting.
AntAngelMed supports a 128K context window.
AntAngelMed achieves inference speeds of over 200 tokens/s on H20 hardware.
AntAngelMed ranks first overall in the MedBench v4 benchmark and leads in 5 dimensions.
AntAngelMed is the best-performing open-source model in the OpenAI HealthBench evaluation, with a particularly clear advantage on the Hard subset.
AntAngelMed is based on the MoE architecture of Ling-flash-2.0 (inclusionAI), with an activation ratio of 1/32.
AntAngelMed's training is divided into three stages: medical-domain continued pre-training, supervised fine-tuning on heterogeneous data (covering mathematics, code and clinical dialogue), and reinforcement learning with GRPO together with a dedicated reward model.
AntAngelMed's BF16 full-precision inference requires 8 Ascend 910B (64GB) cards or 4 Kunlun P800/PPU 810 (96GB) cards.
3 more timeline events
All Facts (10)
AntAngelMed's training data is mainly Chinese and English, and before use in clinical scenarios in other languages, thorough validation is required while keeping scope guardrails and the citation checker activated.
50%UnverifiedAntAngelMed was jointly developed by Ant Health together with the Zhejiang Provincial Health Information Center and Zhejiang Anzhener Medical AI.
50%UnverifiedAntAngelMed uses the Apache-2.0 license, with open weights, allowing commercial use, modification and self-hosting.
50%UnverifiedAntAngelMed supports a 128K context window.
50%UnverifiedAntAngelMed is based on the MoE architecture of Ling-flash-2.0 (inclusionAI), with an activation ratio of 1/32.
50%UnverifiedAntAngelMed's training is divided into three stages: medical-domain continued pre-training, supervised fine-tuning on heterogeneous data (covering mathematics, code and clinical dialogue), and reinforcement learning with GRPO together with a dedicated reward model.
50%UnverifiedAntAngelMed's BF16 full-precision inference requires 8 Ascend 910B (64GB) cards or 4 Kunlun P800/PPU 810 (96GB) cards.
50%UnverifiedAntAngelMed's INT4 quantized version requires only 2 Ascend 910B cards for inference.
50%UnverifiedAntAngelMed is a medical-domain Mixture-of-Experts (MoE) model with 103B total parameters and 6.1B activated parameters.
50%UnverifiedAntAngelMed officially provides an FP8 version, and the community has a GGUF quantized version made by mradermacher that can be used with llama.cpp.
50%