Introduction of Beam: Reflection's 501 Billion Parameter Open-Weight Model

Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, of which 23 billion are active. It is built for coding, reasoning, and agentic workloads.

The model was pretrained on 23.8 trillion diverse tokens from web and proprietary datasets, achieving competitive performance against similar-sized models. It underwent extensive reinforcement learning training, utilizing 10.5K NVIDIA GB300 GPUs over four weeks, generating over 100 million rollouts.

Beam demonstrates strong performance in coding and reasoning tasks, achieving efficiency gains of 3–4 times less inference compute compared to larger models. It is designed to balance capability and token usage, allowing users to adjust reasoning effort based on their needs.

The training involved a large pool of nearly one million environments, focusing on high-quality and challenging tasks. Beam's architecture emphasizes stable optimization dynamics and expert utilization, ensuring effective learning and performance.

The model is currently in final evaluations and will have its weights, technical report, and developer artifacts released soon. Beam's capabilities include advanced reasoning, coding, and tool use, with applications demonstrated in various domains such as software engineering and machine learning workflows.