Reflection releases open-weight model Beam, claiming performance surpasses all open-source models outside China

Reflection releases its first open-weight model Beam, using a MoE architecture that activates 23 billion of 501 billion total parameters per token, focusing on coding, logical reasoning, and agent tasks at low compute. Reflection claims Beam reaches GLM 5.2-level performance on reasoning tasks with 3 to 4 times less compute, with a SWE-bench Verified score of 80.9, but raw performance still trails Kimi K3. Model weights will be released under Apache 2.0 later this month, currently still in final safety testing.