Beam: Reflection's 501B open-weight model
Reflection AI released Beam, a 501-billion-parameter sparse mixture-of-experts open-weight model with 23 billion active parameters, pretrained on 23.8 trillion tokens. The model was built for coding, reasoning, and agentic workloads, with heavy investment in reinforcement learning using 10,500 Nvidia GB300 GPUs over four weeks generating more than 100 million rollouts. Weights, a technical report, and developer artifacts are slated for release later this month.
Why it matters Beam represents one of the largest open-weight models yet released, with a training compute footprint (10.5K GB300s for a month) that rivals closed frontier labs. Its focus on agentic and coding workloads, combined with matching or beating similarly sized open base models, signals that open-weight players are closing the capability gap with proprietary frontier systems rather than just chasing scale for its own sake.
What to watch Watch for independent benchmarks once weights ship later this month, and how Beam's RL-heavy training recipe compares to Mistral's approach with Large 4.