Nemotron 3 Ultra: Nvidia launches open AI model with 550 bi parameters, but China still leads

Nvidia presented Nemotron 3 Ultra, a language model of 550 billion parameters that becomes the best open-weight model in the United States. Despite technical advancement, the system still loses to cutting-edge Chinese models such as Qwen and DeepSeek, highlighting China's leadership in the open AI segment. The announcement was made in Decrypt and reinforces the global race by increasingly powerful and affordable models.

What is Nemotron 3 Ultra and why does this model matter?

Nemotron 3 Ultra is a language model of 550 billion parameters developed by Nvidia with expert mixing architecture (MoE). It was made available with open weights, allowing developers and companies to download, adapt and run it on their own servers — without relying on external APIs. In Nvidia's internal tests, the model surpassed Meta's Llama 3.1 405B and Mistral Large 2, becoming the leader among the US open-weight models.

However, independent benchmarks show that Nemotron 3 Ultra is still behind Chinese models such as Qwen 2.5 72B and DeepSeek-V3 in mathematical reasoning, contextual understanding and computational efficiency tasks. The Chinese advantage is not only in size, but in optimization: Asian country models achieve similar performance with fewer parameters.

Nemotron 3 Ultra vs Chinese models: who leads?

In terms of parameters, Nemotron 3 Ultra is higher than most competitors, but performance does not scale linearly. In the MMLU (general reasoning), GSM8K (mathematics) and HumanEval (code), Qwen 2.5 72B obtained respectively 86.4%, 94.9% and 85.8% of correctness, while Nemotron 3 Ultra recorded 85.1%, 92.3% and 83.5%. The difference is small, but consistent.

In addition, the Chinese DeepSeek-V3 model, with 671 billion parameters, achieved superior performance in energy efficiency and latency, using more refined specialization techniques. This indicates that Chinese leadership is not only crude but strategic: they combine scale with leaner architectures.

Why Nemotron 3 Does Ultra matter to Brazilian developers?

For the Brazilian community of AI, the dispute between US and China opens opportunities. Open-weight models like Nemotron 3 Ultra can be adapted to Portuguese without depending on foreign giants. In addition, competition pushes inference prices and accelerates innovation. Brazilian companies that build AI solutions for agribusiness, health or finance can test Nemotron 3 Ultra in its own Nvidia GPUs, taking advantage of native integration with the CUDA ecosystem.

On the other hand, Chinese models like Qwen already offer multilingual support and can be run on more modest hardware. For budget-limited startups, Chinese models often deliver more cost-benefit. However, privacy and regulation issues — such as the General Data Protection Act (LGPD) — may favor American or European solutions.

Future of open AI: race between USA and China

Nvidia, the world’s largest AI chip maker, has a direct interest in open models running on its hardware — each Nemotron 3 Ultra running on A100 or H100 GPUs generates recurring revenue. Therefore, the company invests heavily in model research, even if it does not market them directly. The announcement of Nemotron 3 Ultra coincides with the expectation for GTC 2026: NVIDIA reveals next-generation AI chips and platform for robotics, where Nvidia should reveal next-generation AI chips that promise to accelerate further the training of massive models.

Meanwhile, China is advancing with state and private investments. The Chinese government subsidizes open AI research as part of its technological sovereignty plan. Models such as DeepSeek-V3 and Qwen 2.5 are developed with support from local universities and big techs such as Alibaba and Baidu, and often surpass Western rivals in open benchmarks. If the trend continues, the US may lose leadership also in the open model segment, which today is dominated by Meta and now by Nvidia.

Frequently Asked Questions about Nemotron 3 Ultra

Nemotron 3 Ultra's free?

Yes, the weights of the model were provided free of charge by Nvidia, but it is necessary to have own infrastructure (GPUs) to run it, which can generate hardware and energy costs.

What are the main differences between Nemotron 3 Ultra and the Chinese model Qwen?

Nemotron 3 Ultra has more parameters (550 bi vs. 72 bi), but Qwen is more efficient and surpasses in reasoning and code benchmarks, besides requiring less computational resources for inference.

How to Test Nemotron 3 Ultra?

Weights are available on Hugging Face and on official Nvidia repositories. It is necessary to have at least one GPU with 80 GB of VRAM to run the full model, or use quantized versions.

Read more about the event GTC 2026: NVIDIA reveals next generation AI chips and platform for robotics and the company's trajectory in the sector. Source: Decrypt.