Nemotron 3 Ultra: Nvidia's Open Model Now Matches Closed AI Systems

Closing the Gap With Closed AI Models

Nvidia said this week that its open-source Nemotron 3 Ultra model now performs on par with the top closed AI systems on enterprise benchmarks — and it does it at one-tenth the cost of running them. That's not a small claim in a market where the price of a single inference call can decide whether an idea ever leaves the whiteboard. The announcement was enough to push Nvidia's shares up roughly 4% the next day, as investors worked through what a cheaper, competitive open model means for everyone else building on top of closed systems.

What Nvidia and LangChain Actually Tested

The numbers came out of a partnership with LangChain, run through LangChain's Deep Agents benchmark suite. Nvidia said Nemotron 3 Ultra reached business task parity with the highest-scoring closed models, while completing more tasks, running at higher throughput, and costing about ten times less per run than those closed alternatives.

Here's the part worth sitting with: none of that came from retraining the model. Every gain came from engineering the system around the model — the tools, the runtime, the agent scaffolding — rather than the model's weights themselves. LangChain cofounder and CEO Harrison Chase framed it plainly: building better agents is less about the model in isolation and more about everything built around it, and this collaboration shows enterprises can get strong performance from an open stack without giving up control of the systems they're running.

Inside the Model

Nemotron 3 Ultra is a 550-billion-parameter Mixture-of-Experts model, with 55 billion parameters active at any given time. It runs on a hybrid Transformer-Mamba architecture and supports context windows up to one million tokens — long enough to hold entire codebases or research threads in a single pass. Nvidia originally released the model on June 4, 2026, built specifically for long-running agentic work: coding agents, deep research tasks, and the kind of multi-step enterprise workflows that fall apart when a model's memory runs out halfway through.

Where Enterprises Are Already Putting It to Work

This isn't staying in the lab. Abridge, Amdocs, and Box are already building specialized agents into their products on top of the model, and EY is expanding its own Nvidia implementation work around what Nvidia calls the NemoClaw blueprint for LangChain Deep Agents. That blueprint pairs the tuned model with Nvidia's OpenShell secure runtime, handing enterprises a fully open stack they can customize and run on their own infrastructure — instead of renting access to something they can't see inside.

Getting Access to Nemotron 3 Ultra

Distribution is already wide. LangChain's agent engineering platform sees more than 200 million monthly downloads, which gives the tuned Nemotron 3 Ultra profile a built-in audience. Beyond that, developers can pull the model through Baseten, Crusoe Cloud, DeepInfra, Fireworks, Nebius, and Together AI.

What It Means for Nvidia's Stock

Shares gained about 4% on Wednesday following the announcement. The move came alongside Bank of America reiterating its Buy rating on the stock on July 8. Even with the bump, Nvidia's stock is only up about 5% year-to-date in 2026, trailing the broader market after a stretch of consolidation — so this is a meaningful data point, not a trend reversal on its own.