Beam Model Technical Report: Architecture and Training of Reflection's Open MoE Model

Reflection has released a technical report for Beam, a sparse Mixture-of-Experts model with 501.3B total parameters and 23.4B active parameters optimized for coding, reasoning, and agentic workflows. Pretrained on 23.8 trillion tokens, the model features a 1-million-token context window developed through a two-stage midtraining phase and a large-scale asynchronous reinforcement learning campaign. Open article →