Reflection Releases 501B MoE Model Beam, Live on Multiple Platforms

Baseten says Reflection's Beam, a 501B open-weight MoE model, is now available on its platform. Beam is built for efficient reasoning in coding and agentic workloads and supports a 1M-token context window. The listing gives developers another open-weight option for long-context coding and agent tasks.

Ollama says it is a day-0 partner for Beam, the first open-weight model from Reflection AI, and is offering it free on its platform for a limited time. The model is hosted in the US with full ZDR, carries a 1M token context window, and is built for software development work.

Artificial Analysis reports that Reflection's first model, Beam, has been released as an open-weights model on Hugging Face with 501B total and 23B active parameters. It scores 30 on the Artificial Analysis Intelligence Index, placing it #2 among US open-weights models.

Reflection released Beam, an open text-in text-out model with a sparse MoE architecture, 501B total and 23B active parameters, and a 1M-token context window. The instruction-tuned model ships in BF16, FP8 and NVFP4, was developed predominantly for English, and the model card highlights token-efficient reasoning that reaches comparable solution quality to leading open models.

Hugging Face's blog introduced Beam, a 501B open model that, according to the post, users can own, customize, and scale. The announcement did not include benchmark results, licensing terms, or the team behind the model.

Reflection's first open-weight model, Beam, is now available on Ollama and free for a limited time. The model targets software engineering, reasoning, and long-running agentic tasks, and can be accessed through Ollama Cloud.

SGLang says it is powering Reflection's Beam model in production. Beam is a 501B-parameter mixture-of-experts model with 23B active parameters per token, built for software engineering and long-horizon agentic tasks, with a planned Apache 2.0 release.