Biomni-R0-32B-Preview

This repo contains the weights of Biomni-R0-32B-Preview, a research preview of the series of biomedical AI agents trained by the Biomni team.

Biomni-R0-Preview is a 32B model trained with end-to-end reinforcement learning using the Biomni-E1 environment scaffolding. It has achieved state-of-the-art performance across ten evaluated biomedical benchmarks spanning diverse tasks including crispr delivery, rare disease diagnosis, gwas variant prioritization, etc.

Read more about how the model is trained and evaluted in our technical report.

Usage: Serve with SGLang

python -m sglang.launch_server --model-path biomni/Biomni-R0-32B-Preview --port 30000 --host 0.0.0.0 --mem-fraction-static 0.8 --tp 2 --trust-remote-code --json-model-override-args '{"rope_scaling":{"rope_type":"yarn","factor":1.0,"original_max_position_embeddings":32768}, "max_position_embeddings": 131072}'

This would require two GPUs with 80G VRAM. Alternatively, you may serve with 4 GPUs with 40G VRAM via --tp 4.

Note, if your task would take significantly longer than the original 32768 context length, you may set rope_scaling factor to a number >1.0 and <=4.0 for smoother context window extension. However, rope_scaling might degrade performance on tasks with shorter trajectories. Please tune the rope scaling factor according to your usage.

To run inference with the Biomni-E1 environment, please follow the instructions in our official repo.

Citation

@misc{biomnir0,
  title     = {Biomni-R0: Using RL to Hill-Climb Biomedical Reasoning Agents to Expert-Level},
  author    = {Ryan Li and Kexin Huang and Shiyi Cao and Yuanhao Qu and Jure Leskovec},
  year      = {2025},
  month     = {September},
  note      = {Technical Report}
}