The latest data from the benchmark testing platform Artificial Analysis shows that Nvidia’s newly launched open source large model Nemotron 3 Ultra is rated as the most powerful open source artificial intelligence model in the United States, but it still lags behind China’s top open source models on an overall level. The model uses a hybrid expert architecture with an overall parameter size of approximately 550 billion, of which approximately 55 billion parameters are actually activated at any time.
In Artificial Analysis' smart score list, Nemotron 3 Ultra scored 48 points, significantly ahead of other US open source models such as Google Gemma 4 31B's 39 points, Nvidia's previous generation Nemotron 3 Super's 36 points, and gpt-oss-120b's 33 points. However, compared with China’s strongest open source model, Nemotron 3 Ultra still lags behind: Kimi K2.6 from Dark Side of the Moon scored 54 points in the same list, while Claude Opus 4.8, the current closed-source model with the strongest overall strength, scored 61 points.
Artificial Analysis's visual chart shows that Nemotron 3 Ultra falls in the "most attractive quadrant" demarcated by it. While its intelligence score is in the first tier, its reasoning and generation speed are also at the forefront. In terms of inference performance, the measured throughput of DeepInfra, the platform hosting Nemotron 3 Ultra, exceeds 300 tokens per second, while DeepSeek and Moonshot models of similar scale can currently only achieve about 50 to 100 tokens per second. Nemotron 3 Ultra has a very obvious advantage in speed.
Nvidia said that Nemotron 3 Ultra will be officially opened to the developer community on June 4, and the first batch will be launched on mainstream model hosting and calling platforms such as Hugging Face and OpenRouter to facilitate integration and testing by enterprises and individual developers. It is generally believed that with the release of this model, the US open source model camp has received an important reinforcement in the high-end capability segment, but there is still a certain gap between China's leading open source model and the international top closed source model in terms of comprehensive intelligence and performance.