AI startup Reflection has released Beam, its first open-weight model designed specifically for coding, reasoning, and agentic tasks. Built to prioritize compute efficiency over raw performance, the mixture-of-experts model activates just 23 billion of its 501 billion total parameters per token. According to the company, Beam matches GLM 5.2 on key benchmarks while consuming three to four times less compute, positioning it as a capable alternative to models built outside China.
The model was developed through a massive training effort that included four weeks of reinforcement learning utilizing 10,500 Nvidia GPUs. Reflection noted that Beam features a tunable parameter allowing users to adjust how thoroughly the model reasons through a problem, effectively trading off compute cost against output quality. Additionally, the text-only model demonstrated emergent capabilities during training, independently learning web browsing behaviors despite lacking direct browsing data in its training mix.
Safety and alignment for Beam are handled by a secondary model merged with the primary architecture, governed by guidelines addressing factual accuracy and hard safety rules. Reflection plans to release the model weights under the Apache 2.0 license alongside a technical report and developer documentation later this month following final safety testing. This launch underscores the startup's ongoing push to provide cost-effective open-weight tools for enterprise coding and automated workflows.