Beam: Reflection's 501B open-weight model

Reflection introduces Beam, a sparse mixture-of-experts open-weight model with 501 billion total parameters and 23 billion active per token, designed for coding, multi-step reasoning, and agentic tool use. Beam was pretrained on 23.8 trillion curated tokens and further scaled via a high-compute reinforcement learning campaign - over 100 million rollouts run on 10,500 NVIDIA GB300 GPUs across four weeks, using up to 256k-token contexts, roughly 1.3 billion sandbox runs, and a million curated environments. Benchmarks show Beam competitive with similarly sized open models and approaching larger families on coding and agentic tasks while delivering 3-4× lower inference compute on advanced reasoning benchmarks, translating into greater “intelligence per token” for enterprise workloads.

Beam’s training emphasizes RL scale and stability: asynchronous policy gradients augmented with new algorithms to control policy staleness and reduce training-inference mismatch, maintaining stable numerics even with samples a day old. A controllable length penalty and a reasoning-effort parameter let users trade response length for performance; early RL improved token efficiency before later expanding reasoning length for harder tasks. Capabilities generalize across domains - browsing, tool use, OCR, and code generation emerged without explicit task presence - and demos include a live NYC subway dashboard, a p5.js astronaut game, a land/sea puzzle with 95.5% coverage, and automated fine-tuning notebook generation. Final red-teaming is ongoing, with weights, technical report, and developer artifacts scheduled for release later this month and early access available by signup.