Reflection Debuts Beam to Challenge Chinese Open Models at Lower Compute

Reflection AI debuted its first foundation model, an open-weight system named Beam, on October 5, 2026, positioning the release to challenge leading Chinese open-weight models at lower compute costs.

Founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, Reflection AI raised over $4 billion and reached a $25 billion valuation prior to the launch. With Beam, the startup is positioning itself as a Western competitor to Chinese releases like Z.ai's GLM series and Alibaba's Qwen series.

Architecture and Training Demands

Beam is structured as a sparse Mixture-of-Experts model containing 501 billion total parameters, with 23 billion active per token. Reflection AI claims this configuration handles coding, reasoning, and agentic tasks while requiring three to four times less compute than rival systems.

According to the company, pretraining drew on 23.8 trillion tokens from curated web content and licensed proprietary datasets. Its post-training phase featured a reinforcement learning run that generated more than 100 million rollouts on 10,500 Nvidia GB300 GPUs across four weeks.

Targeting the Open-Weight Surge

Reflection AI claims Beam matches the performance of Z.ai's GLM-5.2 and approaches Alibaba's Qwen 3.8-Max at a fraction of the token cost. Independent evaluators have not yet replicated or verified these benchmark claims.

The debut coincides with broader developer adoption of accessible model weights. Data from Vercel's AI Gateway cited in reports showed open-weight models accounted for 56 percent of processed tokens in August 2026, rising from 7 percent in December 2025. Reflection AI has backed its expansion with Nvidia investment and government partnerships involving the Pentagon, the U.S. Department of Energy, and international allies including South Korea.

Early Access Questions

Although announced as an open-weight release, Reflection AI launched with an early-access waitlist rather than downloadable model weights or complete technical documentation on debut day. The exact timeline for a general public release of the weights remains unconfirmed.