Physix FrontierPhysix Frontier · News Briefing Card (IT Home · Oct 6, 2026)
Nvidia-backed Reflection AI releases first open-weight model Beam
KEY FACTS
- Reflection AI released its first open-weight model, Beam, on Monday local time.
- The model targets coding and agentic tasks, taking aim at Chinese models such as DeepSeek and Kimi.
- Beam has 501 billion total parameters and activates only 23 billion parameters per task under sparse activation.
- The company says Beam's performance is on par with Zhipu's GLM-5.2, with coding ability approaching Qwen 3.8-Max.
- Reflection AI was founded in 2024 by former DeepMind researchers and previously struck a compute partnership with SpaceX.
KEY DATA
501 billionBeam total parameters
23 billionBeam activated parameters
about 744 billionGLM-5.2 total parameters
40 billionGLM-5.2 activated parameters
PHYSIX OBSERVATION
The US startup uses sparse activation to stack total parameters up to 501 billion while activating only 23 billion per pass, clearly aiming to wrestle with Chinese open-weight models on cost. But its performance only matches GLM-5.2 and approaches Qwen 3.8-Max, showing the gap remains. The open-weight track now has one more Nvidia-affiliated player, and the price war and ecosystem battle will only intensify.
Source: IT Home report