Reflection releases Beam 501B open weights under Apache 2.0

Reflection released the weights of Beam, its first open-weight model, on Hugging Face under an Apache 2.0 license, the company said Friday in a blog post.

What it is: Beam is a sparse mixture-of-experts model with 501 billion total parameters, 23 billion of them active, built for coding, reasoning and long-horizon agentic work, Reflection said. The company published a model card with benchmarks and a safety technical report.

How to run it: Reflection is releasing the weights in BF16, FP8 and NVFP4 formats, plus a speculative decoding drafter model. It says Beam has day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl, TokenSpeed and Nvidia’s Dynamo, and works with coding tools including its own open-source Mirror CLI, Pi, OpenCode and Hermes.

Where to get it: Beyond the open weights and Reflection’s own API, the company says Beam is available at launch through Amazon Web Services, Google Cloud, Microsoft Azure, Oracle, Cloudflare, Databricks, OpenRouter, Ollama, LM Studio and others, part of an ecosystem of more than 40 partners and integrations.

Customizing it: Developers can post-train Beam, including with large-scale reinforcement learning, Reflection said. In its largest training runs it ran roughly 1.3 billion sandboxes across reinforcement learning, evaluation and ablations.

What’s next: Reflection says it plans to release self-distillation recipes for customizing Beam in the coming days.