Reflection AI Releases Apache 2.0 Weights for 501B-Parameter Beam Model

Reflection AI released the open weights for Beam, a 501B-parameter model using 23B active parameters per token under an Apache 2.0 license, positioning it against low-cost Chinese open models. To train the model to solve coding, agent, math, and science tasks in fewer tokens without giving up, Reflection ran more than 100M reinforcement learning rollouts over four weeks on 10,500 Nvidia GB300 GPUs, alongside about 1.3B sandboxes peaking at 883,006 virtual CPUs across 23 clusters on Google Cloud and AWS. The release follows Meta's 816B-parameter Muse Spark Open, which uses 42B active parameters per token and requires a separate commercial agreement for service providers with over $20M in revenue, though neither US model matches top Chinese models on the Terminal-Bench 2.1 command-line agent benchmark according to published scores. US enterprises have widely adopted Chinese open models to cut inference bills—with Coinbase cutting AI spending nearly in half using Z.ai's GLM and Moonshot's Kimi, and DeepSeek handling 17% of Vercel's routed tokens in May 2026 while representing near 1% of developer spend. "What you want to achieve is a model that for its model size uses the minimum number of tokens to achieve the task." — Ioannis Antonoglou, Co-founder and CTO of Reflection AI.