Nvidia hasn’t just been the chip supplier for everyone else for a while now — the company builds its own models too. At Computex 2026, Jensen Huang unveiled Nemotron 3 Ultra, the largest and strongest open model out of Nvidia.
Half a Trillion Parameters — but Sparsely Activated
Nemotron 3 Ultra is the top variant of the Nemotron 3 family (Nano, Super, Ultra). At around 500 billion parameters, the model is huge — but only about 50 billion are activated per token. A hybrid mixture-of-experts architecture makes that possible: instead of running the whole network at every step, only the «experts» needed at the moment switch on. That saves memory and compute. Nvidia promises up to five times higher throughput than earlier versions — thanks in part to its in-house NVFP4 format for the Blackwell platform.
Strong, but Not at the Top
In early independent assessments, Nemotron 3 Ultra is considered the smartest open model out of the U.S. — ahead of other open candidates, but behind the China-led frontier of open models like Kimi K2.6. That’s the honest picture: the U.S. has a first-class open model in Nemotron, but on freely available weights, Chinese labs currently set the pace. On speed, by contrast, Nvidia clearly scores — pre-release tests cite token rates well above comparable open models.
Why Claude Users Should Care Too
Open models are the counterweight to closed flagships like Claude Opus or GPT. They run locally, they’re customizable, and they push prices down across the whole market. Nemotron 3 Ultra clearly targets agentic applications: coding assistants, enterprise search, automated workflows — exactly the fields where Claude and Codex are active too. More serious open competition keeps the closed providers on their toes.
My take: it’s no coincidence that of all players, the chip giant builds one of the strongest open models. The more capable models there are, the more chips are needed — and Nvidia supplies them. An open model here is also a selling point for its own hardware. Cleverly played.
Sources: SiliconANGLE: Nvidia launches Nemotron 3 model family, LLM Stats: AI News