According to CoinMeta, Reflection announced the launch of its first open-weight model, Beam, which adopts a MOE architecture with a total of 501 billion parameters, with 23 billion parameters activated each time. It is mainly aimed at coding, inference, and agent tasks. According to Reflection data, Beam scored 44.4, close to GLM-5.2's 44.0, but it still has not caught up with the latest leading open-source models in China. The training scale of Beam is enormous; pre-training used 6144 NVIDIA GB300, processing 23.8 trillion token, followed by 4 weeks of reinforcement learning using 10,500 GB300 to generate over 100 million training trajectories. Officials claim this is one of the largest reinforcement learning trainings ever publicly recorded. Beam is currently in the security testing phase, and Reflection plans to release the full weight and technical report later this month, and will be made open under the Apache 2.0 protocol.