Reflection AI Launches Beam Open-Weight Model for Enterprises

Reflection AI, a New York City-based startup founded by former Google DeepMind researchers, has unveiled Beam, its first open-weight AI model. Designed as a text-only mixture-of-experts model, Beam targets enterprises, developers, and the public sector for use in reasoning, coding, and agentic tasks. The company asserts that Beam matches the performance of leading Chinese open models, such as Z.ai's GLM-5.2, while operating three to four times more efficiently in token cost and inference compute than competing Western open models.

To support its development, the company secured over $7 billion in compute deals with Space Exploration Technologies Corp. and Nebius Group to access Nvidia GB300 chips through 2029. This includes a $6.3 billion agreement with SpaceX for access to the Colossus 2 data center and a $1 billion deal with Nebius Group.

Beyond the model release, Reflection AI is promoting AI factories that enable institutions to train models on proprietary data. The company is currently testing this sovereign AI concept with South Korea's Shinsegae Group and has previously established partnerships with the U.S. Department of Energy and the U.S. Department of War.