Reflection AI unveils “Beam”, the first open-weight model that challenges the Chinese models

Reflection AI, the American artificial intelligence startup backed by Nvidia, unveiled on Monday its first open-weight model, called Beam — as Reuters reports — aiming to compete directly with Chinese open-weight models such as DeepSeek, Qwen, Kimi and Z.ai on coding tasks and artificial intelligence agents.

Beam is a text-only mixture-of-experts model with 501 billion parameters in total, but activating only 23 billion for each task — according to the company, an architecture that makes it faster and cheaper to run. The model has been trained with 23.8 trillion tokens and has a context window of 1 million tokens, TechCrunch reports. By comparison, the Chinese model Z.ai GLM-5.2 has around 744 billion parameters in total, with 40 billion active.

Reflection claims that Beam matches GLM-5.2 in advanced reasoning tests and outperforms the West’s leading open models while using “3–4 times less inference compute” — but, as TechCrunch stresses, these claims have not yet been independently verified. The company calls Beam a “work model” for enterprises, the public sector and developers, and says it will release the weights and the full technical details this month.

The Brooklyn-based startup, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, has been valued at around 25 billion dollars and has raised approximately 4.7 billion dollars from backers such as Nvidia, Sequoia Capital and Lightspeed. During the summer, Reflection signed deals worth over 7 billion dollars with SpaceX and Nebius to secure access to Nvidia GB300 chips through 2029, according to TechCrunch. The company’s strategy aims at building “artificial intelligence factories” — localised AI systems that allow sovereign institutions and countries to train models on their own data.