Nvidia-backed startup Reflection unveils open-weight AI model said to rival China offerings

Reflection AI, a startup founded by ex-Google DeepMind researchers, has introduced an open-weight AI model named Beam. The company claims Beam competes with leading models from both the United States and China. Beam is engineered for business users and developers, excelling in coding and AI agent tasks while being more cost-effective than many rivals.

According to the firm, Beam achieves results on advanced reasoning benchmarks similar to GLM-5.2, an open model from China's Z.AI. Reflection plans to disclose additional technical specifics about Beam later this month. China has been gradually closing the gap with the US in AI, with Chinese developers like DeepSeek and Moonshot AI gaining traction through affordable and customizable open models.

Leading US entities, primarily OpenAI and Anthropic, have predominantly focused on closed or proprietary models to bolster their enterprise ventures. Reflection emerged from stealth in 2025, aiming to develop open AI models. The startup has secured backing from Nvidia and has engaged in discussions to raise $2.5 billion at a valuation of $25 billion, according to a report by The Wall Street Journal.

Additionally, Reflection has formed billion-dollar agreements with SpaceX and Nebius Group to acquire essential computing resources for model development. This model rollout is expected to strengthen the United States' and allied nations' endeavors to export AI technology abroad. Earlier this year, Reflection was involved in a US-mediated deal to construct sovereign AI in South Korea.

Officials are currently exploring further partnerships to promote the export of US technology, aiming to hinder the proliferation of Chinese alternatives.