IT之家 October 6 news: Reflection AI, a startup backed by Nvidia, on Monday local time launched Beam, its first open-weight large model. The startup hopes to compete with Chinese AI models such as DeepSeek and Kimi in the fields of coding and agentic tasks.
Reflection AI said that Beam's performance is on par with GLM‑5.2 from Chinese AI startup Zhipu AI (Z.ai); in coding and agentic tasks, the model's performance is gradually approaching Qwen 3.8‑Max (Qwen3.8‑Max).
According to the startup, Beam has a total of 501.0 billion parameters; parameter count refers to the variables that determine how an AI system processes information. However, the model uses a sparsely activated architecture, activating only 23.0 billion parameters for each task. By enabling only part of the network each time, it achieves faster inference speed while lowering operating costs.
For comparison, Zhipu AI's GLM‑5.2 has a total of about 744.0 billion parameters, with 40.0 billion activated parameters.
The model's release comes as US tech companies are striving to fend off competitive pressure from Chinese open-weight models. These models are cheaper and more customizable, and their code generation capabilities are nearly catching up with the top models from OpenAI and Anthropic.
According to IT之家, Reflection AI was founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company mainly develops software development automation tools, and software development automation is currently a rapidly growing category of AI applications.
Earlier this year, Reflection AI reached a cooperation agreement with SpaceX, allowing it to access additional computing resources at the Colossus 2 data center of Elon Musk's SpaceX.