Reflection AI has unveiled Beam, an open-weight Mixture-of-Experts model featuring 501 billion total parameters and 23 billion active parameters. Designed for enterprise coding and agentic tasks, the model allows users to adjust reasoning effort to balance task complexity against compute budgets.
The model was pretrained on 23.8 trillion tokens over a period of less than 4 weeks using 6,144 NVIDIA GB300 NVL72 GPUs. The training process achieved a 92.3% goodput rate. Reinforcement learning involved 10,500 NVIDIA GB300 GPUs, generating over 100 million rollouts across 1.3 billion sandboxes.
Beam demonstrates strong performance on industry benchmarks, scoring 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1. The system supports 110,000 concurrent rollouts and updates its inference fleet with new weights in a median time of 12 seconds.
Currently, Beam is undergoing final red-teaming and is not yet available for self-hosting. Interested users can join a waitlist for early access via the Reflection platform. The model is released under permissive Apache 2.0 and MIT licenses.
