Reflection unleashes 501B open-weight AI model, Beam
Reflection has introduced Beam, a 501 billion parameter sparse Mixture-of-Experts model designed for coding, reasoning, and agentic workloads. The model’s capabilities stem from extensive pretraining on 23.8 trillion tokens and high-compute reinforcement learning (RL), generating over 100 million rollouts on 10.5K NVIDIA GPUs. Beam is particularly optimized for efficiency, outperforming larger models like Qwen 3.8-Max on inference compute efficiency while delivering competitive performance on coding and agentic tasks.
The development process involved significant investments in RL science, data, and infrastructure, including new algorithms to maintain stable learning despite policy staleness. Beam’s training also focused on efficient reasoning, allowing users to control the tradeoff between capability and token usage. The model demonstrates strong generalization, applying learned capabilities to tasks beyond its training data, such as web browsing and tool use. Beam is set for release later this month, with early access sign-ups now open.
Read the full article
Continue reading at Hacker News →This is an AI-generated summary. Read the original for the full story.