Reflection debuts Beam, open model rivaling GLM-5.2 on less compute

Reflection AI is a two-year-old, Brooklyn-based startup founded by two former Google DeepMind researchers. It has raised roughly $4.7 billion from backers including Nvidia, Sequoia Capital, and Lightspeed Venture Partners, and its last round valued it at a $25 billion pre-money valuation. Beam is the company's first frontier model, and its launch confirms reporting from Axios over the weekend that a release was close.

The company has been locking up compute at scale, signing deals collectively worth more than $7 billion with SpaceX and Nebius this summer to secure access to Nvidia's GB300 chips through 2029. That compute is meant to support models that can draw customers away from closed labs like Anthropic and OpenAI, as well as the cheaper open-weight models from Chinese labs. Beam itself is a mixture-of-experts model pretrained on 23.8 trillion tokens with a 1 million token context window, trained on high-compute reinforcement learning for reasoning, coding, and agentic tasks.

Beyond shipping Beam, Reflection is aiming future models at enterprises and sovereign nations through an "AI factories" pitch, which would let institutions train Reflection's models on their own proprietary data. Nvidia CEO Jensen Huang has championed the AI factory idea, and Axios reported that hedge funds and trading firms are eager to build such systems. Reflection has already begun testing the concept with Shinsegae Group in South Korea. The stakes for Beam likely hinge on whether its performance claims, which have not been independently verified, hold up once the weights and technical details are released.