100B / 6.1B active (MoE, 1/32 activation ratio)
Apache 2.0
Large open-source medical language model based on Ling-flash-2.0, jointly developed by healthcare and industry partners for professional medical use cases. Official release materials position the model as top-performing on major medical benchmarks, including #1 among open-source models on HealthBench and leading results on MedBench and MedAIBench, with emphasis on medical knowledge QA, diagnostic reasoning, and safety/ethics performance. Training is described as a three-stage pipeline: continual pretraining on medical corpora, supervised fine-tuning with high-quality medical instructions, and GRPO-based reinforcement learning for robust clinical reasoning behavior. Architecture uses high-efficiency sparse MoE design with 100B total parameters and about 6.1B active parameters (1/32 activation ratio), targeting strong capability at lower inference cost; release text reports over 200 tokens/s on H20 hardware and 128K context support. The project also publishes FP8 quantized variants and reports EAGLE3-based acceleration gains under concurrent inference workloads, with specific throughput improvements on HumanEval, GSM8K, and Math-500 settings. Distribution includes Hugging Face, ModelScope, project GitHub, and a public API demo endpoint.