Reflection Beam 1.0

Reflection Beam 1.0 is an Apache-2.0 open-weight sparse mixture-of-experts model for software engineering, tool use, and multi-step agent tasks. Reflection's official Hugging Face card and AWS's October 16 launch documentation describe 501 billion total parameters, 23 billion active parameters per token, and an effective one-million-token context window. The weights are published in BF16, FP8, and NVFP4 variants.

Deployment facts

Large sparse model: Active parameters affect per-token work, but the complete checkpoint, memory sharding, and serving stack must still accommodate 501B total parameters.
Open deployment choice: The launch documents self-managed vLLM and SGLang paths; AWS also lists a SageMaker JumpStart route in US East (N. Virginia) and US East (Ohio) at launch.
Agent-oriented scope: Official descriptions position Beam for resolving software issues, editing code, invoking tools, reading intermediate results, and longer workflows.
Weights, not a hosted price promise: Apache-2.0 covers the posted weights. Infrastructure, API, and managed-endpoint costs must be checked with the provider you choose.

Evaluation workflow

Start with the official model card and a contained coding benchmark or issue-reproduction task. Give the model read-only tools first, cap context and wall-clock time, log tool calls, and compare success, recovery, and cost against a smaller model on the same harness. If using a managed endpoint, confirm the enabled region, quota, retention, and networking policy before sending proprietary repositories. A one-million-token limit is a capacity claim, not a recommendation to fill every request.

Limits and alternatives

Beam's model scale makes local evaluation impractical for many teams even though only 23B parameters are active per token. Context, quantization, parallelism, tool permissions, and benchmark choice can change practical cost and reliability. For a smaller local agent model, compare LFM2.5-24B-A2B; for an existing open-weight coding-focused model, compare Muse Spark Open.