Reflection introduces Beam, a 501B open-weight model

Reflection has introduced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters per token. The company says it trained the model on 23.8 trillion tokens, then used reinforcement learning across more than 100 million rollouts.

The training run used 10,500 NVIDIA GB300 GPUs for four weeks, according to Reflection. The company says Beam is designed for coding, reasoning, and agentic tasks, and that it completed a final red-team review before the announcement.

Beam’s weights, technical report, and model card are planned for later this month. Reflection describes the model as competitive with larger open models, but that comparison is a company claim until independent evaluations are available. This matters because the release could give builders a large open model with a relatively small active parameter count.