Reflection released Beam under Apache 2.0, with approximately 501 billion total parameters, 23 billion active parameters, and 1-million-token context.

Google Cloud’s announcement describes Beam as Reflection’s first open-weight model, released under the permissive Apache 2.0 license. The model has approximately 501 billion total parameters, activates approximately 23 billion parameters per token, and supports an approximately one-million-token context window.

Launch checkpoints are available in BF16, FP8, and NVFP4 numerical formats. The stated model focus is text-native software engineering, complex reasoning and coding, multi-step tool use, and long-horizon agentic tasks.

The listed integrations include Reflection’s Mirror command-line interface, Claude Code, Hermes, and OpenCode. The article identifies resolving GitHub issues and executing command-line and development-and-operations workflows as supported task areas.

The announcement claims Beam reaches solutions using 3–4 times fewer reasoning tokens than comparable open-weight models and describes 3–4 times higher reasoning-compute efficiency. These are attributed model-level claims, separate from Google Cloud’s serving-stack throughput and latency benchmarks; the excerpt does not provide the complete evaluation methodology.