REFLECTION INTRODUCES BEAM MODEL

Summary from the TLDR AI newsletter

Beam is a 501B-parameter sparse Mixture-of-Experts model with 23B active parameters trained for coding, reasoning, and agentic workloads. The model used 23.8T pretraining tokens and more than 100M reinforcement-learning rollouts, with open weights and developer artifacts planned for release later in October.