Reflection AI has launched its first internally trained artificial-intelligence model, escalating the competition to build a U.S.-developed alternative to increasingly capable Chinese open-weight systems.
The model, called Beam, uses a sparse mixture-of-experts architecture containing 501 billion total parameters while activating about 23 billion for an individual task. Reflection says the architecture is designed to reduce computing requirements during inference while maintaining strong performance in coding, reasoning and autonomous-agent workloads.
The launch matters because open-weight AI has become an increasingly important competitive front between U.S. and Chinese developers.
Chinese systems have attracted users partly because businesses can deploy and customize them without relying exclusively on closed proprietary platforms. Reflection says Beam is competitive with Z.ai’s GLM-5.2 and is approaching Alibaba’s Qwen3.8-Max on some coding and agentic tasks. Those comparisons remain company-reported performance claims rather than independently established rankings.
The scale of Beam’s development illustrates how expensive the open-model competition has become. Reflection says the model was pretrained on 23.8 trillion tokens. Its reinforcement-learning program generated more than 100 million rollouts using roughly 10,500 Nvidia GB300 GPUs over four weeks.
Beam is still undergoing final red-team testing and evaluations, meaning the broader public release of its model weights and technical materials has not yet been completed.
That distinction matters for developers and investors. Monday’s announcement establishes the architecture and performance claims, but real-world adoption will depend on availability, operating costs, independent benchmarking and how easily businesses can deploy the system.
For the broader AI market, Beam represents another attempt to challenge the idea that the most competitive customizable models must come from China.
If U.S. developers can deliver high-performing open-weight systems requiring less inference computing, competition could intensify not only among model developers but across the enormous computing infrastructure supporting them.
