Reflection AI unveiled Beam on Monday, the Nvidia-backed startup's first frontier open-weight model. Founded in 2024 by former DeepMind researchers Misha Laskin, the company's CEO, and Ioannis Antonoglou, the startup is pitching Beam as a U.S.-built alternative to cheaper Chinese open-weight models.
The company says Beam packs 501 billion total parameters with 23 billion activated per token, and claims reasoning scores comparable to Z.ai's GLM-5.2 while using three to four times less inference compute for coding and agentic tasks. But the startup's own published results show newer Chinese models — GLM-5.3 and Moonshot's Kimi K3 — still ahead on raw capability, and the benchmarks have not yet been independently verified.
Open weights are due later this month under the Apache 2.0 license, after final red-teaming is complete; an early version is currently available only via waitlist. The launch follows a string of open-model releases from Chinese labs that have pressured Western startups to prove efficiency, not just scale — and Reflection signed a major compute deal with SpaceX's Colossus 2 data center earlier this year to power the effort.
