NVIDIA CEO Jensen Huang unveiled Nemotron 3 Ultra during his keynote at Computex in Taipei on June 1. This open-weight model scored 48 on the Intelligence Index in an AI Analysis and NVIDIA joint pre-release evaluation, but it lost to China’s Moonshot AI’s Kimi K2.6, which scored 54.
Nemotron 3 Ultra Technical Specifications: 55 billion active parameters, million-token context window, and shipping on June 4
Nemotron 3 Ultra uses a Mixture of Experts (MoE) architecture: a total of 550 billion parameters, with only 55 billion active parameters activated at any given time, combined with Mamba-2 layers, standard Transformer attention mechanisms, and mixture-of-experts routing. The model supports a 1 million token context window and uses Multi-Token Prediction (MTP) technology to improve generation speed.
NVIDIA said that, compared with similar open-weight models, Ultra is 5x faster in inference and 30% cheaper. On the DeepInfra pre-release endpoints, the model can process more than 300 output tokens per second; by comparison, DeepSeek V4 Pro and Kimi K2.6 process only 50 to 100 tokens per second via commercial APIs. The model weights are public, the training方案 has been released, and the official shipping date is June 4, 2026.
Direct Comparison of Open-Source AI Intelligence Index: United States vs China — Artificial Analysis Data
Based on evaluation data published by Artificial Analysis, the intelligence index rankings for major models are as follows: all global closed-source flagship models (Anthropic, Google, OpenAI) scored 57. Kimi K2.6 (Moonshot AI, China, released April 2026) scored 54, ranking fourth globally. Nemotron 3 Ultra (NVIDIA, United States) scored 48, ranking first among US open-source models. Google Gemma 4 31B (United States) scored 39. Nemotron 3 Super (NVIDIA, March 2026, 120B parameters) scored 36. OpenAI gpt-oss-120b (United States) scored 33. The Intelligence Index is an aggregated benchmark across 10 evaluation items (inference, coding, general knowledge, and agent performance), and higher scores indicate stronger performance.
The Nemotron Alliance and Nemotron 4: Confirmed Next-Gen Development Framework
At Computex in Taipei, NVIDIA also announced in parallel that Nemotron 4 is already under development. It will be handled by the Nemotron Alliance formed by NVIDIA in March 2026, with eight AI labs including Mistral AI and Perplexity as members, jointly developing based on DGX cloud infrastructure.
NVIDIA previously disclosed a five-year $2.6 billion open-source AI investment plan, and Nemotron 3 Ultra is the most representative milestone to date. The global usage share of China’s open-source models has grown from about 1.2% at the end of 2024 to about 30% at the end of 2025 (source: Decrypt, March 2026 report).
FAQ
What improvements does Nemotron 3 Ultra have compared with the previous-generation Nemotron 3 Super?
Nemotron 3 Super was released in March 2026, with 120 billion parameters and an Intelligence Index of 36. Nemotron 3 Ultra has an Intelligence Index of 48, a year-over-year increase of 12 points. NVIDIA said that a 12-point improvement is a significant leap in the benchmark testing domain.
Why is Kimi K2.6’s Intelligence Index higher than NVIDIA Nemotron 3 Ultra?
China’s Moonshot AI’s Kimi K2.6 has an Intelligence Index of 54, which is 6 points higher than Nemotron 3 Ultra’s 48. It ranks fourth among all models worldwide (including closed-source), trailing only the flagship closed-source models from Anthropic, Google, and OpenAI (each 57) by 3 points. US labs like OpenAI, Anthropic, and Google tend to keep their strongest models behind APIs, while Chinese labs continue to contribute high-scoring models to the open-source ecosystem.
In which scenarios is Nemotron 3 Ultra’s speed advantage most critical?
Ultra processes more than 300 output tokens per second—about 3 to 6 times faster than the commercial API speeds of DeepSeek V4 Pro and Kimi K2.6. NVIDIA said this advantage is especially important when autonomous AI agents execute long-duration, multi-step tasks, because per-step waiting time quickly accumulates in complex tasks.
