A new AI model from Reflection targets coding and software agents while using only a fraction of its parameters at a time. Beam was introduced on October 5, 2026. Early access requires signing up, and downloadable weights are planned for later this month.
The attraction is control over deployment once those files arrive. Weights are the numerical values needed to run a trained model. Until their release, organizations cannot treat the announcement as a ready-to-install alternative to a hosted AI service.
Reflection describes Beam as a text-only mixture-of-experts model. That design routes work through selected parts of a larger system. It has 501 billion parameters in total, with 23 billion active for each token, a small unit of processed text.
The company says its main advantage is efficiency when answering, rather than the highest score on every test. Its own results make that distinction visible. On DeepSWE v1.1, a coding benchmark, Beam scores 44.4, compared with 61.0 for GLM 5.3 and 68.0 for Kimi K3.
Those are Reflection's published comparisons, not a test conducted by this publication. They support a narrower reading of the launch: a new option focused on the cost of running coding and agent workloads.
An agent is software that uses a model to choose actions and call tools. Reflection says Beam's training emphasized this work alongside code and reasoning. The company plans to publish a technical report and model card with the weights, providing more detail about training and limitations.
The launch currently offers early access and developer-reported results. The download is still pending.
