Foreign media: U.S.-based startup Reflection AI has released Beam, an open-weight model aiming to challenge Chinese companies' dominance in the open artificial intelligence space. The model is primarily designed for code generation, reasoning, and intelligent agent tasks. The company claims Beam’s performance approaches that of Zhipu AI’s GLM-5.2 and matches Alibaba’s Qwen3.8-Max in programming and agent capabilities, though it still lags behind Moonshot AI’s Kimi K3 in overall performance.
Beam employs a "mixture-of-experts" (MoE) architecture with a total parameter count of 501 billion, but only activates 23 billion parameters per task, significantly reducing computational demands. The company states that Beam’s inference computation requirements are three to four times lower than those of comparable open-source models, potentially enabling users to run AI applications at substantially lower costs. Preliminary assessments by independent testing organizations suggest Beam may be among the most efficient open-source models available at its level of intelligence.
Original source: toutiao.com/article/1878305059647500/
Disclaimer: The views expressed in this article are solely those of the author.