Nvidia has introduced its most powerful open-weight AI model to date, the Nemotron 3 Ultra, aiming to bolster the U.S. position in the rapidly evolving artificial intelligence landscape. Unveiled at Computex, this 550-billion-parameter model represents a significant leap in performance, boasting speeds that far outpace current Chinese rivals in the open-weight sector. However, despite its advancements, it still trails behind the leading proprietary models emerging from China.
- Nemotron 3 Ultra Launched: Nvidia revealed its largest open-weight AI model, Nemotron 3 Ultra, at Computex on June 1.
- Performance Metrics: The model features 550 billion total parameters with 55 billion active parameters, utilizing a mixture-of-experts architecture for efficiency.
- Speed Advantage: It achieves over 300 tokens per second, significantly faster than comparable Chinese models.
- Intelligence Ranking: Nemotron 3 Ultra scored 48 on the Artificial Analysis Intelligence Index, making it the top U.S. open-weight model, though Moonshot AI’s Kimi K2.6 (score 54) remains the overall leader.
- Technological Foundation: The Nemotron family integrates Mamba-2 layers and Transformer attention, supporting a 1-million-token context window and multi-token prediction.
The Nemotron 3 Ultra utilizes a “mixture-of-experts” (MoE) architecture, a design that allows specific parts of the neural network to be activated for different tasks, akin to a hospital where only relevant specialists are consulted. This approach significantly enhances efficiency, allowing the model to deliver superior inference speeds and reduced operational costs compared to traditional dense models of similar scale. Nvidia claims this results in performance three to six times faster than competing open-weight models from China, with estimated costs being 30% lower.
Independent evaluation by Artificial Analysis placed Nemotron 3 Ultra at a score of 48 on its Intelligence Index, a comprehensive benchmark measuring reasoning, coding, general knowledge, and agentic capabilities. This positions it as the leading U.S. open-weight model, surpassing Google’s Gemma 4 31B (39) and its own predecessor, Nemotron 3 Super (36). The substantial jump from the previous generation underscores Nvidia’s rapid development pace.
NVIDIA just announced the release of Nemotron 3 Ultra in Jensen Huang’s Computex keynote: at 550B parameters (55B active), this is the largest Nemotron 3 model to date, and it is the most intelligent US open weights model
We partnered with @nvidia to evaluate this model for…
— Artificial Analysis (@ArtificialAnlys) June 1, 2026
The Nemotron series, which began in late 2023, now includes variants like Nano for lighter tasks, Super for mid-range enterprise uses, and Ultra for complex reasoning. A key innovation in the Nemotron 3 generation is the incorporation of Mamba-2 layers alongside standard Transformer attention. Mamba-2 is designed for efficient processing of extended sequences, which is crucial for models handling vast amounts of data, such as the Nemotron 3 Ultra’s 1-million-token context window. This capability theoretically allows an AI agent to process entire codebases or extensive research documents simultaneously.
Furthermore, the Ultra model incorporates multi-token prediction (MTP), enabling it to generate multiple future tokens at once, thereby accelerating output generation. Post-training via reinforcement learning across diverse interactive environments has equipped these models with enhanced planning and multi-step task execution abilities. Nvidia is releasing the weights and training methodologies for Nemotron 3 Ultra, making its advanced capabilities accessible, though requiring significant computational resources typically found in data centers.
While Nemotron 3 Ultra offers impressive speed metrics, processing over 300 tokens per second, the intelligence contest remains keenly contested. Moonshot AI’s Kimi K2.6, with a score of 54 on the Intelligence Index, still leads the pack of open-weight models. This gap highlights the ongoing competitive dynamic, where Chinese AI labs have been prolific in releasing high-performing open-source models, significantly increasing their global usage share. Nvidia’s substantial investment in open-weight AI development, including its five-year, $26 billion plan, signals a strong commitment to challenging this trend.
The development of Nemotron 4 is already underway through the Nemotron Coalition, a collaboration involving eight AI labs including Mistral AI and Perplexity, aimed at co-developing cutting-edge open models on Nvidia’s DGX Cloud infrastructure. The release of Nemotron 3 Ultra on June 4 marks a significant milestone in this effort.
Long-Term Technological Impact on Blockchain and Web3
The advancements showcased by Nvidia’s Nemotron 3 Ultra, particularly its enhanced processing speed, massive context window, and efficient architecture, carry significant implications for the future of blockchain and Web3 development. The ability of such sophisticated AI models to process and understand vast amounts of data quickly and cost-effectively can accelerate the development of more intelligent and capable decentralized applications (dApps). For instance, AI agents powered by Nemotron could more effectively analyze complex smart contract code for vulnerabilities, optimize decentralized finance (DeFi) strategies, or provide sophisticated AI-driven interfaces for Web3 gaming and metaverse experiences.
The integration of advanced AI like MoE models into the Web3 ecosystem could also revolutionize Layer 2 scaling solutions. Faster AI processing could enable more efficient state compression, faster transaction validation, and smarter data management on L2 networks, thereby improving scalability and reducing transaction costs. Furthermore, the open-weight nature of models like Nemotron 3 Ultra aligns with the decentralized ethos of Web3, fostering innovation by allowing developers worldwide to build upon and refine these powerful tools. As AI becomes more integrated into the infrastructure of the internet, its synergy with blockchain technology will likely unlock new paradigms in decentralized intelligence, data ownership, and autonomous systems, pushing the boundaries of what is possible in the digital realm.
Details can be found on the website : decrypt.co