SXF VERIFIED BRIEF
Reflection's Beam 1.0 now has Apache 2.0 open weights available in BF16, FP8 and NVFP4 formats. Its published model card lists 501.3B total and 23.4B active parameters, text-only input/output and 1M-token context. That is more specific than the launch post's earlier promise to publish the weights.
Teams investigating self-hosted coding and agents can examine published checkpoints, licensing and multiple numeric formats rather than depending solely on hosted API access. The 23.4B active-parameter figure describes activation per token, not the complete weight memory requirement. Reflection's own Terminal-Bench 2.1 table reports Beam 80.1 versus GLM 5.2 at 81.0 and GLM 5.3 at 88.2; this is one task-specific, vendor-reported comparison, not an independent global ranking.
Use the exact BF16, FP8 or NVFP4 checkpoint and test GPU memory, latency and long-context performance in the target serving stack. Reflection's 3–4x inference-compute claims are estimates excluding prefill and serving overhead; they are not measured production prices. No direct prior Beam release is supplied for a version-over-version comparison, and the model card labels the technical report as forthcoming.