[AINews] Reflection Beam - 501B-A23B American Open Model

Reflection AI has launched its new model, Beam, a 501‑billion‑parameter architecture with 23 billion active weights, designed for programming, autonomous agents, and scientific work. The model was trained from scratch in the United States, and its full weights will be released under an Apache 2.0 license during the month. According to the company, the pre‑training phase consumed 23.8 trillion tokens, some of which came from OCR applied to hundreds of millions of PDF documents. The team also describes a reinforcement‑learning and policy‑optimization stage that used 10,000 GB300 GPUs, producing over 100 million rollouts distributed across roughly one million distinct tasks. In published results, Beam scored 80.9 on the SWE‑bench Verified test and claims to be three to four times more inference‑efficient than comparable models such as GLM 5.2. This initiative represents one of the first large‑scale offerings trained entirely on U.S. soil, expanding the range of options available to developers and companies seeking local AI solutions. Beam’s emergence is significant because it diversifies the open‑model ecosystem and could spur greater competition in performance and cost within the AI market.

Read the original article on Latent Space

This summary is an informational synthesis produced by dataqbs.com. All rights to the original content belong to its author and the cited media outlet. We act solely as curators of technology news and claim no authorship.