Nvidia’s Nemotron 3 Ultra tops US open AI models but trails Chinese rivals

Nvidia’s new Nemotron 3 Ultra has claimed the top spot among open AI models from the United States. Maximilian Schreiner reports for The Decoder that the model scores 48 points on the Artificial Analysis intelligence ranking, putting it clearly ahead of other open US models.

For comparison, Google’s Gemma 4 31B scores 39 points, Nvidia’s own Nemotron 3 Super reaches 36, and gpt-oss-120b manages 33. Nemotron 3 Ultra has roughly 550 billion total parameters, with about 55 billion active at any time.

China’s open models still lead

Despite its strong US ranking, Nemotron 3 Ultra does not reach the performance of the top open models from China. Moonshot’s Kimi K2.6 scores 54 points on the same scale. The current strongest closed model, Opus 4.8, scores 61 points.

One area where Nemotron 3 Ultra stands out is speed. According to Artificial Analysis, the model delivers more than 300 tokens per second on the provider DeepInfra. Comparable models from DeepSeek or Moonshot currently reach only 50 to 100 tokens per second. A token roughly corresponds to a word or word fragment.

Artificial Analysis places Nemotron 3 Ultra in what it calls the “most attractive quadrant” of its evaluation chart, a zone that combines high intelligence scores with fast output speed.

Nvidia plans to release the model on June 4 on Hugging Face, OpenRouter, and other platforms.