Reflection AI Beam benchmarks show 501B open weight model nearing China’s best

Beam is significant not because it clearly beats the strongest Chinese models, but because it meaningfully closes the gap with an Apache 2.0 license. Reflection is openly targeting the “Western open weight frontier,” and Capital & Compute’s breakdown confirms that Beam can compete with GLM 5.2 on many agentic and long context benchmarks while using far less active compute per token. That combination of efficiency and permissive licensing is exactly what enterprises and governments have been waiting for in order to run powerful models on their own hardware.

The training numbers are also a milestone: more than 10,000 top end GPUs and 100 million RL rollouts for a single open weight model signal that serious reinforcement learning at scale is no longer limited to the biggest closed labs. If those tactics diffuse into the broader open source ecosystem once weights and model cards ship, we could see a new wave of highly specialized agentic models tuned for coding, operations and scientific workflows.

For the AGI race, Beam underscores that the competitive frontier is no longer just “closed US labs vs open Chinese models.” We now have Western startups trying to match China’s best on capability while offering legal terms and provenance guarantees that governments in Europe and North America may prefer. That will put additional pressure on giants like OpenAI, Anthropic and Google to decide how far they are willing to go in opening weights or offering strong on prem options.