Reflection finally submits its answer: the 501B model Beam makes its debut, still lagging behind the latest top Chinese open-source models.

动察 Beating AI News Flash: Nvidia-backed AI startup Reflection has unveiled its first open-weight model, Beam. It uses a MoE architecture with 501 billion total parameters, activating only 23 billion at a time, and is primarily aimed at code, reasoning, and Agent tasks.

According to results published by Reflection itself, Beam's overall performance is close to GLM-5.2 and has not yet caught up with the latest top Chinese open-source models. On DeepSWE, Beam scores 44.4, GLM-5.2 scores 44.0, and Qwen 3.8-Max scores 51.0, while GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash reach 61.0, 68.0, and 74.2, respectively.

Beam's training scale is also very large. Pretraining used 6,144 Nvidia GB300 chips to process 23.8 trillion tokens in less than 4 weeks; it was then followed by 4 consecutive weeks of reinforcement learning using 10,500 GB300 chips, generating more than 100 million training rollouts. The company says this is one of the largest reinforcement learning training runs in publicly known records.

Beam is currently still in the final safety testing stage, with only a small number of users able to try it early. Reflection plans to release the full weights, technical report, and model card later this month, and to open them under the Apache 2.0 license.