Reflection AI unveils Beam open-weight model

Reflection AI has unveiled Beam, its first open-weight artificial intelligence model, positioning it for coding, reasoning and tasks in which AI systems plan and act with some autonomy. The startup says the model has 501 billion total parameters, with 23 billion active for each token, and was pretrained on 23.8 trillion tokens. Its weights and technical materials are planned for release under an Apache 2.0 licence later in October 2026. Reflection says Beam is competitive with Z.ai’s GLM5.2 and approaches Alibaba’s Qwen3.8Max on coding and agentic tasks. The company also says Beam matches GLM5.2 on advanced reasoning benchmarks while using about three to four times less inference compute. It describes that efficiency figure as an estimate, not a measurement of serving costs. The claims have not been independently verified in the supplied reporting, and external evaluation will depend in part on the planned release of the model weights. The company acknowledges that Moonshot AI’s Kimi K3 remains stronger in raw capability. Reflection is presenting Beam’s efficiency and performance in selected areas as its competitive case, rather than claiming it leads across all measures.