Reflection AI previews 501B-parameter Beam model

Reflection AI unveiled Beam, previewing its first open-weight model for coding, reasoning and agentic workloads. The company says Beam is a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active parameters per token, trained on 23.8 trillion tokens.

Reflection also disclosed unusually large reinforcement-learning infrastructure for the model. It says a four-week run used 10,500 Nvidia GB300 GPUs and generated more than 100 million rollouts across coding, agentic and STEM environments. The company published benchmark results showing the model competitive with several larger open-weight systems, though those results are self-reported and a full technical report is still pending.

Beam is currently in final red-teaming and evaluation. Reflection says it plans to publish the weights, model card, technical report and developer artifacts later in October under an Apache 2.0 licence, making this an early technical preview rather than a completed public release.