Reflection AI has unveiled Beam, its first open-weight model featuring 501 billion total parameters, with 23 billion active parameters per request. Built on a Mixture of Experts architecture, the model is designed to optimize inference costs while maintaining performance levels comparable to Chinese counterparts like GLM-5.2 and Qwen3.8-Max in coding and agentic tasks.
The training process for Beam utilized Nvidia infrastructure, involving 10 500 NVIDIA GB300 GPUs for reinforcement learning over four weeks. The base model was trained on 23.8 trillion tokens using a cluster of 6144 NVIDIA GB300 NVL72 units. During development, the system supported up to 110 000 parallel runs with a context window reaching 256 000 tokens.
Reflection AI claims that Beam achieves reasoning performance on par with GLM-5.2 while consuming 3–4 times fewer computational resources during inference. The company currently provides early access to a limited group of users for safety testing. A full release, including model weights under the Apache 2.0 license, technical documentation, and developer tools, is scheduled for this month.
