Reflection AI releases Beam, an efficient open-weight model

Reflection AI, a Brooklyn-based startup, released Beam, an open-weight text-only mixture-of-experts model with 501 billion total parameters and 23 billion active parameters. The model is claimed to be on par with GLM 5.2, which has 744 billion total and 40 billion active parameters, while using three to four times less inference compute. Beam scored 44.4 on the DeepSWE software engineering benchmark, compared to 68 for Kimi K3 and 74.2 for DeepSeek V4.1 Flash. The training process involved asynchronous reinforcement learning to manage staleness across 100 million rollouts, with gaps of up to 107 model versions between updates. The team noted that returns to compute investment were still strong at the end of the RL phase, implying further capability gains are possible. The White House has finalized an exemption sparing open-weight models from pre-release security reviews, a policy described as politically risky as model capabilities improve.