US lab Reflection releases Beam, a 501-billion-parameter open-weight model under Apache 2.0

The US AI lab Reflection on 16 October 2026 released Beam, its first open-weight model, publishing the weights on Hugging Face under the Apache 2.0 licence, which allows commercial use and modification.

What the company says

• Beam is a sparse mixture-of-experts model with 501 billion total parameters, of which 23 billion are active, built for coding, reasoning and long-horizon agent tasks.
• Weights come in BF16, FP8 and NVFP4 formats, plus a speculative-decoding “drafter” model to speed up generation. A safety technical report and a model card with benchmarks were released alongside.
• Reflection also open-sourced Mirror CLI, a terminal coding harness with a Rust core that it trained Beam alongside.
• The company says it launched with more than 40 partners: the model is available via cloud providers including AWS, Google Cloud, Microsoft Foundry and Cloudflare, inference services such as Together AI, Fireworks and OpenRouter, and local tools such as Ollama and LM Studio; Dell and HPE offer on-premises and air-gapped deployments.
• During its largest reinforcement-learning runs, Reflection says it used roughly 1.3 billion sandboxes.

Why it matters

Beam joins a busy week of large open models, a day after Meta’s open-weight Muse Spark. An Apache 2.0 licence is more permissive than the custom licences used for many big open models.

All capability and benchmark claims are Reflection’s own; independent evaluations will take time.