Reflection AI Raises the Stakes in Western Open-Weight Models

Reflection AI's Beam is presented as a 501B-total, 23B-active Apache-licensed MoE with million-token context, agentic coding focus, a reported 10,000 GB300 training run, and roughly 100 million RL rollouts across one million tasks. The announcement intensifies U.S.–China open-model competition and highlights reinforcement-learning scale and inference economics. Weights, independent evaluations, access, and company-reported efficiency claims remain unresolved.