AI Consortium Releases Open 26B Biomedical Reasoning Model Trained With Reinforcement Learning, Achieving 81% Drug Evaluation Accuracy

A consortium of AI companies including Loka, AWS, and Arcee AI releases Trinity Mini, an open 26-billion-parameter biomedical reasoning model trained with reinforcement learning that achieves 81.2% drug evaluation accuracy, offering organizations an auditable, deployable solution for specialized scientific AI within controlled infrastructure.

• Loka, AWS, Arcee AI, and Prime Intellect post-train an open 26-billion-parameter mixture-of-experts model called Trinity Mini using two reinforcement-learning environments — Drug Tool RL and BioReason — to teach it evidence-seeking, biological reasoning, and structured scientific output.
• After 21 controlled experiments, Run 120 emerges as the promoted checkpoint, achieving 81.2% on the held-out Drug Tool evaluation and a 0.863 combined score on BioReason, which measures GO term accuracy, tree similarity, aspect coverage, and JSON validity.
• The promoted adapter now runs inside a multi-agent AI Scientist application built with Strands, FastAPI, and React on AWS ECS Fargate, offering organizations a reusable, auditable post-training method for deploying specialized open models within controlled infrastructure boundaries.