The American DeepSeek Arrives After 20 Months: Reflection Releases Beam, a 501-Billion-Parameter Open Model Measured Against Two Chinese Models

On October 5 New York-based Reflection AI released Beam, its first open-weight model: a sparse mixture-of-experts with 501 billion total parameters and 23 billion active per token.

Three key facts

According to Semafor, Reflection says Beam needs three to four times less compute to reason through a problem than comparable open models. It performs on par with Z.ai's GLM-5.2 from June, approaches Alibaba's Qwen 3.8-Max, and trails the top closed models from OpenAI and Anthropic. Its strengths are coding and agentic tasks.

Reflection was founded by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou. Backers include Nvidia, Sequoia and Citigroup, and a June round valued the company at $25 billion pre-money. When it raised $2 billion in October 2025 the pitch was to be America's open frontier lab. The original early-2026 release target slipped all the way to now.

Laskin stated the selling point outright: buyers who want sovereign AI without a Chinese model do not really have very good options today. Reflection is working with the U.S. Center for Advancing Innovation and Standards for Super Intelligence and the UK AI Security Institute on evaluations.

WangDou's Take

Twenty months of waiting for the American DeepSeek, and who is it benchmarked against? Zhipu's GLM-5.2 and Alibaba's Qwen 3.8-Max, two Chinese models. Matches one, approaches the other, trails OpenAI and Anthropic. In plain language: America's top open model gets its report card graded on a Chinese ruler. A $25 billion valuation produced something roughly equal to a Chinese model from June. Laskin's line about buyers having no good options is really about his own positioning: not the strongest, just not Chinese. Nvidia and Citigroup did not invest in a model. They invested in the label alternative. One note from an old founder: open source is won on ecosystem, not parameter count. DeepSeek won by shipping three versions in six months. Beam's first version took 20.

Source: Semafor · Reflection AI unveils an open-source answer to Chinese labs · Reflection AI Blog