Nvidia-backed Reflection AI unveils 501-billion-parameter open model to challenge China

Reflection AI, a two-year-old startup based in Brooklyn, introduced its first model on Sunday: an open-weight system called Beam that the company says can match leading Chinese AI models on reasoning and coding tasks while using far less computing power to run.

Beam is a sparse "mixture of experts" model with 501 billion total parameters, of which 23 billion are active for any given query, and was trained on 23.8 trillion tokens of text, code and technical data. It can process context windows of up to 1 million tokens, according to Reflection's announcement on its blog introducing Beam.

The company said Beam is "competitive with larger open models like GLM-5.2 and approaching Qwen 3.8-Max on coding and agentic tasks," citing internal benchmark scores including 77.2 on SWE-Bench Pro v2-Hard and 97.8 on the AIME 2026 math competition test. Reflection's central pitch is efficiency: it says Beam delivers reasoning performance comparable to the Chinese-made GLM-5.2 while requiring three to four times less inference compute, which would make it markedly cheaper to run at scale. Those comparisons have not been independently verified.

A Western answer to Chinese open models

Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both former Google DeepMind researchers, and has positioned itself explicitly as a domestic counterweight to Chinese open-weight labs such as DeepSeek, Alibaba's Qwen and Z.ai's GLM, which have moved aggressively to give away powerful models for free. The strategy has attracted serious money: Nvidia led an $800 million slice of a $2 billion funding round that valued Reflection at roughly $25 billion before the new money, with backing also from Sequoia Capital and Lightspeed Venture Partners, according to reporting from TechCrunch.

The company has also locked up enormous amounts of computing power to support the effort, including multiyear deals with SpaceX and the cloud provider Nebius for access to Nvidia's GB300 chips. Reflection said Beam itself was pretrained on 6,144 GB300 GPUs in under four weeks, with a subsequent reinforcement-learning phase using more than 10,000 GPUs that generated over 100 million training rollouts, according to Reflection's own technical writeup.

For now, Beam is not fully public. Reflection is offering early access to a limited group of users through a waitlist while the model finishes final evaluations. The company said it plans to release Beam's weights openly under an Apache 2.0 license later this month, along with a technical report, model card and tools for running and fine-tuning it — a rollout aimed at developers, enterprises and government agencies rather than casual chatbot users, according to additional coverage.

The release lands amid growing anxiety in Washington and Silicon Valley that Chinese labs have pulled ahead in the race to give away capable, freely modifiable AI models, which are increasingly used as the foundation for other companies' products. Reflection's launch is the most concrete evidence yet that a well-funded American challenger is prepared to compete directly on that turf, rather than conceding the open-source lane to China. Whether Beam's performance claims hold up once outside researchers and developers can test it directly will determine if Reflection becomes a serious rival to the Chinese labs it is chasing, or simply the latest well-capitalized startup promising a model it has not yet fully shipped.