Meet AntAngelMed: A 103B-Parameter Open-Source Medical Language Model Built on a 1/32 Activation-Ratio MoE Architecture

Researchers in China released AntAngelMed, an open-source medical language model with 103B total parameters and a 1/32 MoE activation ratio.

The model was trained in stages on medical data, instruction tuning, and reinforcement learning to improve clinical reasoning and task performance.

It is designed for efficiency as well as capability, with longer context handling and faster inference for real-world healthcare use.

Reported optimizations include FP8 quantization, speculative decoding, YaRN, and EAGLE3 support on H20 hardware.