U.S. AI startup Reflection AI has unveiled its first large open-weight model, Beam. On Oct. 6 local time, Business Insider reported that Reflection AI is positioning the model as a U.S.-made alternative to China’s powerful open AI systems.

Beam is a text-only model tailored for coding, reasoning and AI agent tasks. Developers can download the model parameters, run it on their own hardware, and modify or fine-tune it for their use cases. That differs from models that are mainly accessed through their own APIs, such as OpenAI or Anthropic.

Reflection AI said Beam has 501 billion parameters in total, with 23 billion activated at a time. The company said this structure lowers operating costs and increases processing speed. Beam is currently undergoing final safety tests and is being provided via a waitlist. Reflection AI plans to release the model weights, a technical report and developer tools within this month.

Reflection AI was founded in 2024 by former Google DeepMind researchers Misha Laskin (미샤 라스킨) and Ioannis Antonoglou (이오아니스 안토노글루). Chief Executive Misha Laskin said in a podcast released on Oct. 6 that the only way to own intelligence is when it is open, adding that it must be open so users can run it on their own equipment, control it and customize it.

Reflection AI said Beam can compete with Z.ai’s GLM-5.2 in coding and agent tasks and is close to Alibaba’s Qwen 3.8-Max. It assessed that in pure performance, Moonshot AI’s Kimi K3 still leads. It also said Beam delivers similar results to GLM-5.2 on advanced reasoning benchmarks while using 3 to 4 times fewer reasoning computations needed for prompt responses.

Reflection AI has received investment from Nvidia, Sequoia and Citigroup, among others. In June, it confirmed a fundraising round at a pre-money valuation of $25 billion. Some investors call Reflection AI the "West’s DeepSeek," and Nvidia’s official AI account on X, along with investors Didi Das and Shaun Maguire, also congratulated the Beam launch.