Reflection AI Unveils Beam: Cheaper, Faster AI Power

Reflection AI has officially launched Beam, its inaugural frontier, open-weight AI model. This move marks a significant challenge to existing leaders in the AI landscape. Beam is engineered as a text-only mixture-of-experts model, optimized through reinforcement learning for superior performance in reasoning, coding, and agentic tasks.

The company claims Beam achieves parity with top-tier Chinese open models on advanced reasoning benchmarks, all while operating at a fraction of the cost and inference time. Beam is a 501-billion-parameter model, featuring 23 billion active parameters and an extensive 1 million token context window. It was pre-trained on 23.8 trillion tokens, positioning it as a powerful tool for enterprises and the public sector.

Founded in 2024 by former Google DeepMind researchers, Reflection AI has secured substantial backing, including from Nvidia. The startup has also strategically committed over $7 billion in deals with SpaceX and Nebius to ensure access to critical compute resources through 2029. This ensures their capacity to develop and deploy advanced AI models.

Reflection AI aims to build 'AI factories,' allowing institutions to create customized, local AI systems using their proprietary data. This vision aligns with Nvidia CEO Jensen Huang's advocacy for open AI ecosystems. The company is already piloting this concept with Shinsegae Group in South Korea, signaling a strong push towards sovereign AI solutions.