Reflection AI announces Beam, a 501-billion-parameter open-weight model

What is changing?

Reflection AI introduced its first open-weight model, Beam, on October 5. Its sparse Mixture-of-Experts (MoE) architecture activates 23 billion out of 501 billion total parameters. It focuses on coding, reasoning and tool-using AI agents.

The company reports results close to GLM-5.2 on some reasoning tests while using three to four times less inference compute. This is an approximate compute estimate that does not include all service costs; it is not a direct price comparison or an independent performance test.

Beam is still undergoing final safety tests and evaluations. Initial access is limited to a waitlist. Reflection plans to publish the weights under the Apache 2.0 license, along with a technical report and developer tools, during October.

Our creative-production assessment: Beam does not directly fit the image or video generation category. Its coding capabilities and ability to automate repetitive production tasks with tools are more relevant. Its practical contribution will need testing once access opens.