Nvidia just dropped its biggest open-weight AI yet — but China still leads the pack.
At Computex in Taipei, CEO Jensen Huang unveiled Nemotron 3 Ultra: a 550-billion-parameter model that actually activates only about 55 billion parameters at any moment thanks to a mixture-of-experts design. Think of it like a hospital with hundreds of specialists where only the relevant doctors show up for each case — it gives headline-scale capability without headline running costs. Nvidia says that design yields roughly 5x faster inference and 30% lower operating costs than comparable open-weight alternatives.
How smart is it? Independent evaluator Artificial Analysis, which partnered with Nvidia on a pre-release assessment, put Nemotron 3 Ultra at 48 on its composite Intelligence Index (a 10-test aggregate covering reasoning, coding, knowledge and agentic tasks). That makes Ultra the smartest U.S. open-weight model to date — ahead of Google’s Gemma 4 31B (39), Nemotron 3 Super (36) and OpenAI’s gpt-oss-120b (33). It’s also a big jump from Nemotron 3 Super (120B, released March 2026): Ultra is 12 index points higher, a meaningful leap in this benchmarking landscape.
Architecture and capabilities
- Nemotron 3 family (first launched Nov 2023; third generation announced Dec 2025) ships in Nano, Super and Ultra sizes. All use a hybrid architecture that mixes Mamba-2 layers (an alternative attention mechanism optimized for very long contexts), standard Transformer attention, and mixture-of-experts routing.
- Ultra supports a 1-million-token context window — enough to hold a large codebase or hundreds of documents in memory at once — and uses multi-token prediction (MTP) to accelerate generation by predicting several tokens at once.
- All Nemotron 3 models received post-training via reinforcement learning across interactive environments, improving planning and multi-step execution. Nvidia is releasing Ultra’s weights and training recipes publicly.
Speed vs. raw intelligence
Ultra shines on throughput: on a pre-release DeepInfra endpoint it served over 300 output tokens per second, compared with 50–100 tokens/s currently typical for Chinese commercial APIs like DeepSeek V4 Pro and Kimi K2.6. That speed matters for latency-sensitive deployments such as autonomous agents or high-frequency pipelines.
But speed isn’t the whole story. China’s Kimi K2.6 (Moonshot AI) scored 54 on Artificial Analysis’s index and ranks fourth globally among all models (open or closed), just three points behind proprietary flagships from Anthropic, Google and OpenAI (all at 57). In short: Nvidia has closed the gap for U.S. open-weight models, but China remains ahead in raw measured intelligence.
The geopolitical open-weight race
Chinese labs have aggressively pushed strong open models into the ecosystem — open-model usage jumped from roughly 1.2% of global usage in late 2024 to about 30% by the end of 2025. American companies have generally kept their top systems behind APIs, so Nvidia’s public wager is notable: a disclosed five-year plan to spend $26 billion on open-weight AI development, with Nemotron 3 Ultra the most visible result so far.
What this means for crypto and Web3
For crypto-native projects, Nemotron 3 Ultra has practical implications:
- Faster model throughput and large-context windows enable more capable on-chain/off-chain agents, richer smart-contract analysis, and larger-scale on-chain data indexing and audit tooling.
- Public weights and recipes lower the barrier for decentralized projects and researchers to build, fine-tune and audit models for tokenized data, DAO governance assistants, or automated trading strategies.
- The datacenter-class hardware requirements mean smaller teams will still access Ultra via Nvidia’s API or cloud providers rather than running it locally — similar to how most projects use GPT or Claude today — which keeps centralized cloud providers in the loop.
What’s next
Nvidia is already planning Nemotron 4, developed through the Nemotron Coalition — eight AI labs including Mistral AI and Perplexity — built on DGX Cloud infrastructure. Nemotron 3 Ultra ships June 4.
Bottom line: Nemotron 3 Ultra is a major milestone for U.S. open-weight AI — faster and more capable than prior domestic options and openly published — but it doesn’t yet overtake China’s top open models on measured intelligence. For crypto builders, its public weights, huge context window and speed promise new tooling and agent possibilities, even if big-model compute will remain a datacenter play for now.
Read more AI-generated news on: undefined/news
