Reflection released Beam 1.0's weights on October 16, turning its earlier announcement into a downloadable model that developers can self-host and customize. Beam uses a sparse mixture-of-experts architecture with 501.3 billion total and 23.4 billion active parameters per token, a one-million-token context and BF16, FP8 and NVFP4 variants. The text model targets software engineering, tool use, math and science reasoning and long-horizon agent work. Apache 2.0 weights arrive with deployment containers, inference integrations, customization pipelines and more than 40 ecosystem partners. The release expands self-hosting and domain adaptation, but the roughly one-terabyte BF16 repository, compute cost, evaluation, safeguards, English emphasis and June 2026 knowledge cutoff remain practical limits.