AntAngelMed: A 103B Total / 6B Active Medical MoE Model

The dilemma facing medical AI gets cracked open by MoE architecture

Healthcare institutions deploying AI have long faced a frustrating trade-off: large models are reliable enough, but their inference costs are prohibitively high; small models are affordable to run, but they keep making mistakes on critical judgments. AntAngelMed's answer takes a middle path—using a Mixture-of-Experts architecture to pack the knowledge capacity of 103B parameters into the inference overhead of 6.1B activated parameters.

This model, jointly developed by Ant Health, the Zhejiang Provincial Health Information Center, and Zhejiang Anzhen'er Medical AI, is released under the Apache-2.0 license and supports a 128K context window. It achieves inference speeds exceeding 200 tokens/s on H20 hardware. For hospitals and health tech companies that need localized deployment, this means clinical triage, medical record summarization, and decision support can run on their own infrastructure, without patient data ever leaving the security boundary.

Mixture-of-Experts (MoE) is a sparsely activated neural network architecture. Traditional dense models invoke all parameters when processing each token, whereas MoE splits the feed-forward layers into multiple "expert" sub-networks, with a learnable routing gate network deciding that each token activates only a few of these experts. This allows the model to have an extremely large total parameter count (i.e., knowledge capacity), while the parameters actually involved in computation per inference (activated parameters) are far fewer. AntAngelMed's 103B total parameters / 6.1B activated parameters, an activation ratio of about 1/32, is precisely a manifestation of this mechanism: achieving knowledge reserves far beyond its scale with compute overhead close to a 6B dense model. The cost of MoE lies in the fact that VRAM usage is still calculated based on total parameters, and imbalanced routing loads may cause training instability, so techniques such as load balancing loss are often needed.

Benchmarks: comprehensive leadership in Chinese medical scenarios

On MedBench v4—currently the most rigorous benchmark in the Chinese medical field (with physician participation in evaluation)—AntAngelMed took first place overall, leading in 5 dimensions. In OpenAI's HealthBench evaluation, it is the best-performing open-source model, with a particularly notable advantage on the Hard subset (high-difficulty clinical cases).

AntAngelMed's strengths are concentrated in medical question answering and ethical safety dimensions. The latter is precisely a direction many models tend to overlook—medical AI not only needs accurate knowledge, but also needs to respond in ways that conform to ethical norms in boundary scenarios.

Architecture and training: polished in three stages

AntAngelMed is based on the MoE architecture of Ling-flash-2.0 (inclusionAI), with an activation ratio of 1/32. Training proceeds in three steps: first, continued pretraining in the medical domain; then supervised fine-tuning on heterogeneous data (covering mathematics, code, and clinical dialogue); finally, reinforcement learning using GRPO with a dedicated reward model.

Regarding hardware requirements, BF16 full-precision inference requires 8 Ascend 910B (64GB) cards or 4 Kunlun P800/PPU 810 (96GB) cards; after INT4 quantization, this drops to 2 Ascend 910B cards. An official FP8 version is provided, and the community also has a GGUF quantized version produced by mradermacher available for use with llama.cpp.

On language support, special attention is needed: the training data is primarily Chinese and English. If you want to use it for clinical scenarios in languages such as Portuguese or Spanish, be sure to conduct thorough validation first, while keeping scope guardrails and citation verifiers active.

GRPO (Group Relative Policy Optimization) is a policy optimization algorithm commonly used in the reinforcement learning alignment stage, and can be viewed as a simplified variant of PPO. It does not rely on an independent value network (critic); instead, it samples a group of multiple responses for the same prompt and uses the mean reward within the group as a baseline to calculate relative advantages, thereby greatly reducing VRAM and training costs. In medical scenarios, combined with a dedicated reward model, GRPO can impose preference constraints on dimensions such as answer accuracy, citation standards, and ethical safety, making model outputs closer to clinically acceptable behavioral boundaries. Compared with supervised fine-tuning, which can only imitate standard answers, the reinforcement learning stage enables the model to learn "which answer is better" in open-ended, multi-path clinical dialogues. This is also one of the important reasons why AntAngelMed performs outstandingly in the ethical safety dimension.

The Ascend 910B and Kunlun P800/PPU 810 mentioned in the article are both domestic AI accelerator cards: the former comes from Huawei's Ascend series, with 64GB HBM per card; the latter comes from Baidu Kunlunxin, with 96GB per card. BF16 full-precision inference requires 8 910B cards or 4 P800 cards. The main bottleneck is not compute power but VRAM—although the MoE model activates only 6.1B parameters, all 103B weights still need to reside in VRAM. INT4 quantization compresses weights to 4-bit integers, reducing VRAM requirements to about one quarter, so only 2 910B cards are needed. The official FP8 version strikes a compromise between precision loss and VRAM savings. GGUF is a quantization format used in the llama.cpp ecosystem, suitable for offline operation on consumer-grade or edge devices. The community version provides another deployment path for institutions lacking high-end accelerator cards.

Regulatory red lines: ranking first on benchmarks does not mean you can practice medicine

No matter how good the benchmark scores are, they cannot authorize anyone to practice medicine. Taking Brazil as an example, diagnoses must be made by physicians registered with the CFM (Law No. 12.842/2013), telemedicine is governed by CFM Resolution No. 2.314/2022, software with medical purposes falls under ANVISA's RDC 657/2022, and medical records are sensitive personal data (LGPD Article 11).

A gateway layer like g.cloud has its core value in ensuring that the output of AntAngelMed (or any model) passes through a safety gate before reaching humans: no diagnosis without a doctor, no leakage of sensitive data, and every decision has a publicly auditable record.

The Brazilian regulations listed in the article have general reference significance: CFM is the Brazilian Federal Council of Medicine, responsible for physician registration and practice standards; ANVISA is roughly Brazil's drug and health regulatory agency, and RDC 657/2022 brings software with medical purposes under medical device regulation; LGPD is Brazil's General Data Protection Law, and Article 11 classifies health data as sensitive personal data, requiring stricter legal bases for processing. This combination reveals a common pattern: compliance for medical AI involves at least three independent regulatory lines—professional qualifications, software market access, and data protection—and none is exempted just because a model scores high. Other jurisdictions (such as the EU MDR, the U.S. FDA SaMD framework, and China's NMPA-related regulations) follow similar logic. Before deployment, local regulations should be checked item by item rather than relying on a single benchmark score.

Frequently asked questions

Can it be used commercially? Yes. Apache-2.0 license, open weights, modifications and self-hosting allowed.

How to integrate? Serve it through vLLM or SGLang with an OpenAI-compatible API. Gateway-type products can connect as proxies, SDKs, or plugins without modifying the model itself.

Is it suitable for non-Chinese and non-English markets? It needs validation first. Training focuses on Chinese and English, and clinical use in other languages must undergo independent evaluation.

Key facts

• MoE architecture: 103B total parameters / 6.1B activated parameters, Apache-2.0, 128K context
• First overall on MedBench v4, best open-source model on HealthBench (notable advantage on the Hard subset)
• Inference speed over 200 tokens/s on H20, official FP8 version + community GGUF quantization
• Developers: Ant Health + Zhejiang Provincial Health Information Center + Zhejiang Anzhen'er Medical AI

One-line Summary

AntAngelMed uses MoE architecture to achieve 103B knowledge capacity with 6.1B activated parameters, topping Chinese medical benchmarks.

Paragraph Summary

AntAngelMed is a medical large model jointly developed by Ant Health, the Zhejiang Provincial Health Information Center, and Zhejiang Anzhen'er Medical AI. It uses a MoE architecture with 103B total parameters but only 6.1B activated, achieves inference speeds over 200 tokens/s on H20, supports 128K context, and releases open weights under the Apache-2.0 license. It ranks first overall on MedBench v4 and becomes the best-performing open-source model in OpenAI HealthBench, particularly excelling in high-difficulty clinical cases and the ethical safety dimension. Training is divided into three stages: continued medical pretraining, supervised fine-tuning on heterogeneous data, and GRPO reinforcement learning. On hardware, BF16 requires 8 Ascend 910B cards or 4 Kunlun P800 cards, while INT4 quantization reduces this to 2. The article also emphasizes that benchmark scores do not equal medical practice qualifications, and deployment must comply with local medical regulation and data protection laws.