Reflection announces Beam with an open-weight release planned for October
October 5. Reflection announced Beam, a sparse mixture-of-experts model aimed at coding, reasoning and agentic tasks. It has 501B total parameters and activates 23B per token. Reflection describes performance as competitive with GLM-5.2 and approaching Qwen3.8-Max; these are the developer’s comparisons, not an independently established ranking. Full weights and technical materials are planned for later in October. Reflection’s announcement.
Active parameters are only part of the hosting question. The full weights still need to be accommodated in the deployment, alongside runtime memory and the workload’s context needs. Hardware requirements, release licensing and task-level results will matter when assessing Beam as a self-hosted coding-agent foundation.
Sponsored
ChatGPT Codex
OpenAI's agentic coding tool that runs in ChatGPT, your IDE, and the terminal to handle complex engineering tasks like bug fixes, refactors, and PR reviews.
View tool
Comments
No comments yet
Be the first to share your thoughts