Reflection AI Announces Beam, a 501B Parameter Open-Weight Model
Reflection AI has introduced Beam, its first open-weight language model featuring 501 billion total parameters and 23 billion active parameters in a Mixture-of-Experts (MoE) architecture. Released under the Apache 2.0 license, Beam is tailored for coding, reasoning, and agentic workloads and is currently in final red-teaming. Beam expands the available options for massive open-weight models, offering a practical open-source alternative for high-level reasoning and agentic tasks. By activating only a fraction of its parameters during inference, it promises to significantly reduce computational costs while serving complex AI applications. Reflection AI claims Beam requires 3 to 4 times less inference compute than models like GLM-5.2 while maintaining benchmark-comparable reasoning performance. An early version of the model is currently available through a waitlist ahead of its full release.
## BACKGROUND
An open-weight AI model makes its pre-trained parameter weights publicly accessible, allowing users to download, run, and fine-tune the model on their own infrastructure. A Mixture-of-Experts (MoE) architecture optimizes resource usage by dynamically routing input tokens to specific sub-networks rather than processing every token through all 501 billion parameters.