Beam previews a coding model; downloadable weights are still pending

Reflection’s Beam is a mixture-of-experts model aimed at coding, reasoning and tool-using agents. It has 501 billion parameters, with 23 billion active for each token. The company describes extensive reinforcement learning across executable tasks and sandboxes. Access currently goes to selected early users; the promised weights and supporting technical artifacts have not yet been released.

Reflection · Oct 5, 2026

Why it matters

Teams considering self-hosted coding models should distinguish a preview announcement from a model they can deploy today.

More context

Watch the early-access program and the promised release later this month. Reflection plans Apache 2.0 weights plus running, evaluation and fine-tuning documentation. Its training account explains how asynchronous rollouts, model-version tracking and sandbox orchestration support long tool-using tasks.

Keep in mind

Beam remains under final evaluation and red-teaming. Capability and infrastructure numbers are vendor reports, and 23 billion active parameters do not establish a 23-billion-parameter storage requirement. We have not tested it.

Original read