NVIDIA's Nemotron 3 Ultra tops US open models but lags behind Chinese rivals in intelligence
550B parameter open model beats Chinese in speed, but Kimi K2.6 leads in intelligence scores
NVIDIA used its GTC Taipei 2026 keynote to announce Nemotron 3 Ultra, its most powerful open-source AI model developed in the US. With 550 billion parameters, the model is set to be released open-source within the first week of June 2026. Benchmarks show it scores 48 on the Artificial Analysis Intelligence Index, outperforming Google's Gemma 4 31B (39) but falling short of Chinese open models like Kimi K2.6 (54). However, Nemotron 3 Ultra compensates with significantly faster token generation speed, making it highly cost-effective for inference-heavy applications.
Alongside the model, CEO Jensen Huang confirmed mass production of the Vera Rubin AI server, a data center system combining the Rubin GPU and Vera CPU, delivering 2400 TFLOPS in FP64 precision for agentic AI. NVIDIA also introduced the RTX Spark SoC (Arm CPU + GPU for laptops) and the DGX Station Windows desktop with up to 748GB memory, capable of running models up to 1 trillion parameters. These launches solidify NVIDIA's strategy of providing both open models and infrastructure for enterprise AI deployment.
- Nemotron 3 Ultra has 550 billion parameters and scores 48 on the AI Intelligence Index, best among US open models.
- Model loses to Chinese rival Kimi K2.6 (54 points) but offers higher output speed and cost-efficiency.
- Vera Rubin AI server enters mass production with 2400 TFLOPS FP64, targeting agentic AI factories.
Why It Matters
US open-source AI competitiveness hinges on speed and cost, but Chinese models still lead in pure intelligence scores.