Reflection AI boosts US ambition to compete with Chinese ‘open’ models

Reflection AI, a US start-up backed by Nvidia and the Trump administration, unveiled its first open-weight model, Beam, claiming on Monday that it is on a par with Z.ai’s GLM-5.2, a leading Chinese model, and outperforms US peers on coding and agentic tasks “at a fraction of the token cost”. Founded in 2024, it has raised more than $4bn and reached US$16.2B valuation before releasing a model; Nvidia has invested $800mn and provided chip access, and other backers include Sequoia, Disruptive and 1789 Capital (where Donald Trump Jr is a partner). It has partnered with the Pentagon and the Department of Energy and has agreements to build models for US allies including South Korea. Demand context: open-weight models were 56% of tokens processed through Vercel’s AI Gateway in August, up from 7% in December. Policy context: Treasury secretary Scott Bessent has floated sanctions on Chinese models accused of IP theft and urged more US open-weight models to avoid “regulatory capture” by OpenAI and Anthropic; an Nvidia-led industry letter defended an open-weight ecosystem, Jensen Huang has defended US firms’ use of Chinese open models for value, and Republican lawmakers are probing national-security risks of dependence on PRC-jurisdiction models. Reflection says Beam could intensify a price war and will launch soon with partners. Caveats: the parity claim is the company’s own, full technical details have not been published, and the FT does not independently verify benchmarks.