Biomni-R0-32B-Preview: A Biomedical AI Agent

Biomni-R0-32B-Preview is a 32 billion parameter research preview model developed by the Biomni team at Stanford. This model represents a significant step in biomedical AI, focusing on the development of AI agents capable of complex reasoning within the biomedical domain.

Key Capabilities & Training

• Reinforcement Learning: The model is trained using end-to-end reinforcement learning, leveraging the specialized Biomni-E1 environment scaffolding.
• Biomedical Task Specialization: It is specifically designed and optimized for a range of biomedical tasks.
• State-of-the-Art Performance: Biomni-R0-32B-Preview has demonstrated state-of-the-art performance across ten evaluated biomedical benchmarks. These benchmarks cover diverse applications such as CRISPR delivery, rare disease diagnosis, and GWAS variant prioritization.
• Context Length: The model supports a substantial context length of 32768 tokens, with flexibility for extension up to 131072 tokens using rope scaling, though this may impact performance on shorter tasks.

Usage and Further Information

This model is intended for research and development of biomedical AI agents. Detailed information on its training and evaluation methodology can be found in the technical report. For running inference within the Biomni-E1 environment, instructions are available in the official repository.