Reflection AI releases open-weights Beam model to rival DeepSeek and Kimi, IT Home reports
Reported by IT Home · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Reflection AI, an Nvidia-backed startup, released Beam, its first open-weights large model with 501B total parameters and 23B active parameters, IT Home reported. The company says Beam matches Zhipu AI's GLM-5.2 and approaches Qwen3.8-Max on coding and agent tasks.
China context
- Outside China
- Open weights · ithome.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers outside China can evaluate Beam as an open-weights alternative for coding and agent tasks, with claimed performance near Qwen3.8-Max and lower inference cost due to sparse activation.
- For investors
- Nvidia's backing and the SpaceX compute deal signal capital and infrastructure support for a US startup challenging Chinese open-weights models in the developer tools market.
Translated from Chinese. Quotes and facts link to the original sources.
Reflection AI, founded in 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, launched Beam, an open-weights sparse model with 501B total parameters and 23B active parameters. The company claims Beam's performance is comparable to Zhipu AI's GLM-5.2 (744B total, 40B active) and is approaching Qwen3.8-Max in coding and agent tasks. Reflection AI has a partnership with SpaceX to access additional compute at the Colossus 2 data center.
Beam uses a sparse activation architecture, activating only 23B of its 501B parameters per task, which the company says enables faster inference and lower running costs. For comparison, GLM-5.2 has 744B total and 40B active parameters.
US AI companies face increased competition from lower-cost, customizable Chinese open-weights models whose coding capabilities nearly match those of OpenAI and Anthropic; Reflection AI's Beam is a direct response aimed at the developer and agent market.
Beam targets developers building coding and agent applications, offering an open-weights alternative to Chinese models like DeepSeek, Kimi, GLM-5.2, and Qwen3.8-Max, with potential cost advantages from sparse activation.
Independent benchmarks and real-world developer adoption will show whether Beam's sparse architecture delivers competitive coding and agent performance at lower cost. The SpaceX compute partnership may support further scaling.