Beam: Reflection's 501B open-weight model
Summary
Reflection AI unveils Beam, its first open-weight 501B sparse Mixture-of-Experts model, optimized for coding, reasoning, and agentic tasks. The post details large-scale RL training, 23.8T curated tokens, efficiency gains versus frontier models, safety/alignment efforts, and plans to release weights and documentation under an Apache 2.0 license.