NVIDIA-backed startup Reflection AI has unveiled Beam, its first open-weight AI model, designed to challenge dominant Chinese models like Z.ai's GLM-5.2. Valued at $25 billion, the Brooklyn-based startup claims Beam matches rival performance on coding and reasoning benchmarks while using 3 to 4x less inference compute. The launch intensifies the US-China AI race, drawing support from Trump administration officials who advocate for Western open-weight alternatives to counter Chinese dominance.
Technical specifications and training of Beam
- ▪Reflection AI used 10.5K NVIDIA GB300 GPUs for four weeks to run high-compute reinforcement learning for its open-weight model Beam, generating over 100 million rollouts
- ▪Beam is a sparse Mixture-of-Experts model containing 501 billion total parameters and 23 billion active parameters per token
- ▪Beam was pretrained on 23.8 trillion tokens from web, public, and proprietary licensed datasets, with pretraining completed in under four weeks on 6,144 NVIDIA GB300 GPUs
Benchmark performance of Beam
- ▪On Terminal Bench v2.1, Beam scored 80.1, placing it close to GLM-5.2 at 81.0, but behind DeepSeek V4.1 Flash at 90.6 and Kimi K3 at 88.3
- ▪On SWE-bench Verified, Reflection AI's open-weight model Beam scored 80.9 compared to 70.7 for NVIDIA's Nemotron 3 Ultra, according to benchmarks reported by Reflection AI
- ▪Reflection AI claims its open-weight model Beam matches the performance of Z.ai's GLM-5.2 on reasoning benchmarks while using three to four times less inference compute
Efforts to counter Chinese AI competition
- ▪Reflection AI is positioning Beam to compete directly with leading Chinese open-weight models, including Z.ai's GLM-5.2, DeepSeek, and Moonshot AI's Kimi K3
- ▪Reflection AI has partnered with the Pentagon, the US Department of Energy, and South Korea to develop AI models to counter Chinese competition
US debate over Chinese AI models
- ▪US Treasury Secretary Scott Bessent and other Trump administration officials have floated imposing sanctions on Chinese AI models over intellectual property theft concerns
- ▪NVIDIA CEO Jensen Huang has defended the use of Chinese open-weight models by US businesses, stating they provide good value for companies
Securing chip access
- ▪Reflection AI secured access to NVIDIA GB300 chips through 2029 by signing deals worth over $7 billion with SpaceX and Nebius, including capacity at SpaceX's Colossus 2 data center
- ▪NVIDIA has invested $800 million into Reflection AI and provided access to NVIDIA's chips to help the startup compete with Chinese rivals
Debatable claims
- ▪Highly efficient, smaller AI models can successfully compete with larger frontier models
- ▪Open-weight AI models are more effective at driving innovation than closed models
Story comments
Loading comments…