Reflection Beam is Reflection's first open-weight model, a 501-billion-parameter mixture-of-experts model aimed at coding, reasoning and agent work.
In plain words
Beam is a mixture-of-experts model: of its 501 billion parameters, only 23 billion are active at a time. Reflection pretrained it on 23.8 trillion tokens, then trained it with reinforcement learning on 10.5 thousand NVIDIA GB300 GPUs for four weeks, with a focus on coding and agents. Its developer docs list a context window of 1M tokens. On 16 October 2026 Reflection published the weights on Hugging Face under the Apache 2.0 licence, so anyone can download the model.
A quick example
Imagine a bank that cannot send its code to an outside service. With the weights public, it could run Beam on its own servers as the brain of a coding agent.
Where you'll meet it
- Coding and agent tasks, the focus of its training.
- Apps that need reasoning, tool calling and structured output through Reflection's API.
Related
Sources
- DocsIntroducing Beam: Reflection's 501B open-weight model, Reflection · read 5 Oct 2026
- DocsModels, Reflection Developer Docs · read 16 Oct 2026
- DocsBeam is now available, Reflection · read 16 Oct 2026
Checked by WhatIsAI against the sources above · 16 Oct 2026. A full five-minute explainer with a diagram may follow. Spotted an error? Report it.