Nvidia Just Dropped Nemotron 3 Ultra: 550B Parameters of Pure Agent Chaos

At a keynote in Taipei, Jensen Huang pulled the curtain back on Nemotron 3 Ultra, the absolute heavyweight of their open-weights Nemotron 3 family. This beast clocks in at roughly 550 billion parameters, making it a direct rival to the likes of DeepSeek. Nvidia isn't just dropping the weights on Hugging Face and ModelScope; they are practically begging developers to use it for building autonomous AI agents.

Under the hood, the architecture is a clever hybrid of Mamba-Transformer layers paired with latent Mixture of Experts. While the model boasts 550 billion parameters, it only triggers about 55 billion per token, keeping things surprisingly nimble. Thanks to the Mamba integration, it handles a massive 1 million token context window with linear complexity. Nvidia claims it is up to 5 times faster and 30% cheaper than its peers, though these internal benchmarks are about as objective as a parent describing their child's athletic prowess.

The model shows raw muscle in instruction following and long-context handling, even if it occasionally trips over itself in complex coding tasks or long-term planning compared to specialized models like Kimi or Qwen. Beyond just the weights, Nvidia is dumping about 3 trillion tokens of pre-training and post-training data, alongside their NeMo Gym and NeMo RL libraries. They clearly want to be the infrastructure provider, not just a model shop.

Perplexity is already routing tasks through the Nemotron agent router, signaling that the corporate race for agentic supremacy is effectively turning into a server-selling scheme. Whether this open-source push is a genuine gift to humanity or just a clever way to ensure that every AI agent in existence eventually relies on Blackwell architecture remains the real question. It turns out, giving away the blueprint is the easiest way to make sure the world builds the house exactly how you want it.

Source: Coaley Peak