tl;dr: Reflection is releasing Beam, a 501B-parameter (23B active) open-weight Mixture-of-Experts model targeted at coding, reasoning, and agentic tasks, pretrained on 23.8T tokens and trained with large-scale RL (100M+ rollouts on 10.5K GB300 GPUs over 4 weeks). It claims competitive performance with larger open models like GLM 5.2 while using 3–4× less inference compute, with weights to be released under Apache 2.0 later this month.