Reflection AI Announces First Open Source Model, Beam

Reflection AI, a 2-year old Nvidia-backed startup, announced its first open-weight model, adding to a collection of U.S. competitors aiming to catch up to the performance of similarly free models from China.

In its announcement Monday, Reflection said it is still evaluating and testing the long-awaited model and plans to release it later this month. Reflection aims to be the U.S. answer to Chinese companies like DeepSeek, whose AI models have dominated the open weight market. Open weight models are free to download, which offers greater security, customization and predictable pricing compared to proprietary models from companies like OpenAI and Anthropic.

Beam has about 500 billion total parameters, the numbers that make up an AI model and determine its behavior. That makes it about half the size of Thinking Machines Lab’s open weight model Inkling and around the same size as Nvidia’s Nemotron 3 Ultra, which the chip giant released in June. Reflection says that Beam outperforms both models in tests of coding and reasoning capabilities, but comes in behind leading Chinese models.

More western companies have turned to developing open weight models in recent months. Nvidia is also working on a larger Nemotron model, with assistance from a coalition of AI companies that includes Reflection.