Reflection releases open-weight model Beam, rivaling DeepSeek and Qwen with lower compute
Reflection announces its first free open-weight model Beam, positioned to achieve strong performance with minimal compute, rivaling DeepSeek and Qwen's open-weight models. Beam is a MoE model with 501B total parameters, 23B activated per token; according to Reflection, it matches GLM 5.2 on reasoning tasks with three to four times less compute, and approaches Qwen3.8-Max on coding and agent benchmarks, but raw performance still falls short of Kimi K3.
Return to original text
A short introduction is retained here; the full text from the source website is not reposted. For specific wording and subsequent updates, please refer to the original article.
Open AIHOT original