Reflection AI, a startup founded in 2024 by two former Google DeepMind researchers, has released an open-weight model called Beam, built for corporate users and code developers, particularly for creating and running AI agents.
The company says Beam performs as well as the leading Chinese open models, such as DeepSeek and Moonshot, while costing a fraction of what Anthropic and OpenAI charge for their large language models. Reflection has not yet published the model's weights or full technical details but says it will do so soon.
Reflection's backers include Nvidia and Sequoia. The startup has just closed a new funding round at a pre-money valuation of $25 billion; the amount raised was not disclosed.
Open-weight models publish their architecture and parameters so other developers can download and adapt them, unlike closed-weight models such as ChatGPT and Claude, which are accessible only through their providers' cloud or API. According to the report, open-weight models now account for close to 80% of tokens processed industry-wide, a reversal from the start of the year, when closed models held that 80% share.
France's Mistral has made a similar move, launching Mistral Large 4 (also called "le Chonk"), designed for efficiency in cybersecurity, programming and finance, with the same pitch: matching the best Chinese open models at a fraction of the cost of closed American ones. Mistral, described in the report as Europe's leading AI startup, raised €3 billion in a Series D round in September at a €21 billion valuation, according to CNBC.