Reflection AI Announces Beam Model
Nvidia-backed startup Reflection AI has announced Beam, its new 501-billion-parameter model, aimed at competing with low-cost Chinese rivals.
Nvidia-backed startup Reflection AI has officially unveiled Beam, its first open-weight model developed to rival Chinese artificial intelligence systems. Featuring a text-based mixture-of-experts architecture, the model matches the performance of leading Chinese systems in advanced reasoning tests while consuming less inference compute power.
Technical Specifications of the Beam Model
Developed by Reflection AI, Beam is a text-based mixture-of-experts model boasting a total of 501 billion parameters and 23 billion active parameters.
The new model was designed to compete with low-cost Chinese artificial intelligence systems and deliver superior performance in coding and agent-based tasks.
Performance Comparison with Chinese Competitors
According to the company's statements, Beam matches the performance of leading open-source Chinese models, such as Z.ai's GLM-5.2 model, in advanced reasoning benchmarks.
The model achieves this success while using significantly less inference compute power compared to its rivals, providing a cost advantage.
Background of the Startup and Infrastructure Agreements
Founded in 2024 by experts who were former Google DeepMind researchers, Reflection AI rapidly secured major investments and partnerships.
The Nvidia-backed artificial intelligence startup signed strategic computing capacity agreements with major technology players such as SpaceX and Nebius.