NVIDIA ships Nemotron 3 Ultra and Nano Omni — the US open-weights answer arrives the same day
Jensen Huang's Computex keynote unveiled Nemotron 3 Ultra (550B total, 55B active, 90% sparsity) and Nemotron 3 Nano Omni in the same drop. NVIDIA is calling Ultra the most intelligent open-weights model released by a US lab.
NVIDIA used Jensen Huang's Computex 2026 keynote on June 1 to close out the Nemotron 3 family with two simultaneous releases: Nemotron 3 Ultra, a 550B-total / 55B-active sparse model at 90% activation sparsity, and Nemotron 3 Nano Omni, a multimodal perception model that handles document, audio, and video reasoning in a single open checkpoint. Ultra is being positioned as the most intelligent open-weights model released by a US lab to date; Nano Omni is being pitched as the perception sub-agent for the kind of agentic stack NVIDIA wants enterprises to assemble themselves.
The contrast with the same-day MiniMax M3 launch is the part worth watching. Both Nemotron 3 releases shipped with weights on Hugging Face, technical report, and the NVIDIA Open Model License live on day one — the playbook the open-weights community has spent two years standardizing. Nano Omni alone topped six leaderboards for complex document intelligence and audio and video understanding at launch, and NVIDIA published the data recipe and training tooling alongside the checkpoints. That's the difference between a model an engineer can fine-tune on Tuesday morning and a model whose weights are in a press release.
The strategic frame is that Nemotron 3 Ultra is NVIDIA's answer to a question the company doesn't usually have to answer in public: what does a US frontier open-weights stack actually look like when the leaderboard has been dominated by DeepSeek, Qwen, Moonshot, MiniMax, and Z.ai for the past four quarters? The answer NVIDIA is offering enterprises is a US-origin model under a permissive-enough license that you can run it inside a VPC, with a Nano variant in the same family for edge and agentic perception work — and a vendor whose chips you're already buying. That's a different sales motion than "trust our benchmarks for ten days."
NVIDIA Newsroom — NVIDIA Debuts Nemotron 3 Family of Open Models → · Crypto Briefing — Nvidia CEO Jensen Huang launches Nemotron 3 Ultra AI model at Computex 2026 → · Hugging Face — Introducing NVIDIA Nemotron 3 Nano Omni: Long-Context Multimodal Intelligence for Documents, Audio and Video Agents →