Reflection has made Beam’s weights available for coding, reasoning and long-horizon agent workloads. The October 5 preview promised a later release; October 16 marks that availability, confirmed the same day by Ollama.
The MoE model has 501 billion total and 23 billion active parameters. Reflection releases BF16, FP8 and NVFP4 formats, a speculative-decoding drafter and safety documentation. The stated license is Apache 2.0. Integrations include vLLM, SGLang, Transformers and several agent harnesses. Reflection also offers Mirror CLI, with tool approvals, resumable sessions and agents in separate worktrees.
Weights can be downloaded and self-hosted. Reflection’s first-party API is described as a public beta. Ollama provides beam:cloud, free for a limited time with a stated zero-data-retention policy for that endpoint.
beam:cloud is not local inference; active parameters do not represent total memory requirements. Efficiency gains and demos remain vendor evaluations.
Moving from a preview to downloadable weights changes deployment and customization options without creating a second, entirely new model.