On Monday, October 5th, the startup company Reflection AI, backed by Nvidia, launched its first open-source weight model “Beam”. The company aims to compete in the fields of programming and intelligent agents with cheaper Chinese models like DeepSeek and Kimi.
Based in Brooklyn and founded just two years ago, Reflection AI stated that Beam’s performance in advanced reasoning benchmark tests can rival leading Chinese open-source models while significantly reducing costs. This claim is expected to intensify competition between Western models and Chinese counterparts like DeepSeek, Qwen, and Z.ai.
In a detailed blog post on Monday, the company revealed more information about Beam, describing it as a “MoE” (Mixture of Experts) model that supports only text processing. The model, trained using high computational power reinforcement learning, shows excellent performance in reasoning, programming, and intelligent agent tasks, with “token costs and reasoning computation times only a small fraction of competitors”.
With a total of 501 billion parameters, including 230 billion activation parameters, Beam was pre-trained on 23.8 trillion tokens and supports a context window of 1 million tokens. In comparison, Z.ai’s GLM-5.2 model has about 744 billion total parameters with 40 billion activation parameters.
The claims about performance from Reflection AI have not been independently verified yet, but the company stated that in advanced reasoning benchmark tests, Beam’s scores are on par with Z.ai’s GLM-5.2 and surpass current leading Western open-source models, while the reasoning computation is only a quarter to a third of the latter. Reflection positions Beam as a “flagship model” for enterprises, public sectors, and developers.
This release comes at a time when U.S. tech companies are facing competition from Chinese open-source weight models. This summer, Reflection signed agreements worth over $7 billion with SpaceX and Nebius to secure the use of Nvidia’s GB300 chip before 2029.
Reflection plans to introduce Beam and its subsequent models to the enterprise and sovereign state markets. The core concept is to build an “AI factory” by training Reflection’s AI models using an organization’s proprietary data, allowing institutions to establish customized local AI systems. Nvidia’s CEO Jensen Huang has long advocated for the concept of the “AI factory” and has been committed to strengthening the open AI ecosystem. This vision will also benefit Nvidia as these systems will be powered by Nvidia’s GPUs.
Founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, Reflection is dedicated to developing tools that automate software development, which is a rapidly growing application in the AI field.
Earlier this year, Reflection reached an agreement with SpaceX to access additional computational resources through SpaceX’s Colossus 2 data center.