Reflection AI has unveiled Beam, its first open-weight model. Both sources describe it as a 501B-parameter sparse mixture-of-experts model with 23B active parameters, designed for coding and agentic workloads. The company claims it matches Z.ai's GLM-5.2 on advanced reasoning benchmarks while using 3-4x less inference compute.
The two accounts agree on the core specs and performance claim. They differ on licensing: MarkTechPost says the weights are Apache 2.0, while TechCrunch only says open-weight and that weights and technical details will be released this month. TechCrunch adds that Beam is text-only, pretrained on 23.8T tokens, with a 1M token context window.
TechCrunch also stresses that the performance claims haven't been independently verified, and positions Beam against Chinese open models like DeepSeek, Qwen, and Z.ai, as well as Western open models such as Inkling. It notes Reflection's benchmark shows Beam outscores Inkling on four coding tests, but Inkling is multimodal.