Nvidia CEO Jensen Huang unveiled the company's most powerful open-source AI model yet at Computex Taipei on June 1. But benchmark scores show it still lags behind a Chinese rival.
The Nemotron 3 Ultra, a 550-billion-parameter open-source model, scored 48 on the Artificial Analysis intelligence index, while China's Kimi K2.6, released in April, scored 54.
America's best open-source card still hasn't hit China's ceiling.
What Huang Unveiled
The model is Nvidia's most capable open-source offering yet:
- 550 billion total parameters, 55 billion activated — MoE architecture, only a fraction is used per inference
- 1 million token context window
- Over 300 tokens per second output speed
- Claims to be 30% cheaper and 5x faster than comparable rivals
Architecturally, it's a mix: Mamba-2 layers + Transformer attention + MoE routing + multi-token prediction. In short, it combines the speed and cost-reduction techniques of the past two years into one model, aiming for the same intelligence but faster and cheaper.
The model will be available on June 4 on HuggingFace, ModelScope, and OpenRouter, with a NIM microservice wrapper on build.nvidia.com.
Benchmark Breakdown
Nvidia compared Ultra with several models on stage, but according to the third-party intelligence index, the rankings are:
| Model | Developer | Intelligence Score |
|---|---|---|
| Kimi K2.6 | Moonshot (China) | 54 |
| Nemotron 3 Ultra | Nvidia (US) | 48 |
| Gemma 4 31B | 39 | |
| Nemotron 3 Super | Nvidia | 36 |
| GPT-OSS-120B | OpenAI | 33 |
The strongest open-source model on this list comes from China, and it's not a narrow win — 54 vs. 48, a full 6-point gap. Kimi K2.6, released in April, now ranks fourth globally among all models (including closed-source).
Nvidia's two open-source cards score 48 and 36; OpenAI's GPT-OSS-120B trails at 33.
How US Open-Source Fell Behind
The awkwardness isn't about a single score, but the broader landscape.
In China, DeepSeek, Moonshot, Alibaba, and Zhipu AI have been releasing flagship-level capabilities as open-source, available for anyone to download and run. In the US, the most capable models — GPT-5.5, Claude, Gemini — are locked behind paywalls, with only downgraded versions open-sourced.
As a result, Chinese labs lead the open-source intelligence race, with the US chasing.
Nvidia clearly doesn't accept this. It has invested $26 billion in open-source initiatives, with the Nemotron series even releasing training data publicly, aiming to re-establish that "US open-source can compete." Ultra's 48 is the best result yet from this strategy.
But 48 still falls short of 54.
Conclusion
Nvidia, which sells chips to the world for training models, entering the open-source model game itself is a telling signal — it doesn't want to just be the shovel seller; it wants a voice in the model layer.
But its opponents happen to be Chinese teams that, despite using restricted chips, still outperform it in open-source benchmarks.
How long will it take Nvidia to close that 6-point gap? That story will unfold with the next version number.
Sources: Nvidia Releases Its Best Open AI Model Yet—But Still Lags Behind China (Decrypt); Nemotron 3 Ultra announced: high-speed, leading US open weights intelligence (Artificial Analysis); CocoLoop; Nvidia CEO Jensen Huang launches Nemotron 3 Ultra AI model at Computex 2026 (Crypto Briefing); NVIDIA Computex 2026: Complete Recap (explainx.ai)