Reflection releases Beam 1.0 open weights for agentic coding

Reflection released the open weights for Beam 1.0, a sparse MoE positioned for software engineering and long-horizon agent tasks. Official material describes 501 billion total parameters, 23 billion active parameters per token, and an effective one-million-token context window. The weights are published under Apache-2.0 in BF16, FP8, and NVFP4 variants through Reflection's Hugging Face organization.

AWS documented a managed SageMaker JumpStart route at launch in US East (N. Virginia) and US East (Ohio). Teams that need more runtime control can instead deploy with vLLM or SGLang. Neither option changes the planning constraint that the complete checkpoint, memory sharding, and serving topology must accommodate the 501B model, not only the active parameter count.

For coding-agent teams, this creates an open-weight option with a self-controlled deployment boundary. Evaluate it first against a contained issue-reproduction suite with read-only tools, logging context use, latency, tool calls, and recovery from failures. Confirm region, quota, retention, and networking before sending proprietary repositories to a managed endpoint. A one-million-token limit and open weights are not guarantees of low cost, modest hardware requirements, or reliability.

Related catalog entry