Nemotron 3 Ultra Is Live — NVIDIA's Frontier Orchestration Reasoner

June 6, 2026

Nemotron 3 Ultra is in the model picker today. NVIDIA's 550B total / 55B active hybrid Transformer-Mamba MoE is built for long-running agentic workflows — multi-step planning, tool chains, and orchestration that has to hold context across dozens of turns. It ships at Workhorse pricing, is swarm-eligible, and is in the Auto routing pool on every plan.

What you can do

• Pick Nemotron 3 Ultra [Global] from the model selector — routed through OpenRouter for global availability.
• Run long-horizon agentic jobs — research loops, multi-tool engineering, and scheduled workflows where the model keeps a plan across many tool calls.
• 256K context window — enough for a dense brief, a large source set, or a long artifact chain in one session.
• Reasoning with tool support — same agentic surface as other Workhorse picks: search, code execution, documents, presentations, and the rest of your toolbox.
• Use it in Swarm — Nemotron 3 Ultra is swarm-eligible when you want a capable orchestrator or specialist worker without jumping to frontier-tier cost.
• Let Auto pick it — it sits in the Workhorse tier of the auto-router pool, alongside Kimi K2.6 and MiniMax M3, for sessions where breadth and planning matter more than absolute frontier quality.

Where this shows up

• You're running a multi-day project — market research, a product launch pack, or an engineering refactor — and need a model that plans the sequence, calls tools in order, and doesn't lose the thread halfway through.
• You want NVIDIA's largest Nemotron tier for orchestration-style work but don't need to pay SOTA-band prices. Ultra gives you the 550B MoE architecture at Workhorse token rates.
• You're already on Nemotron 3 Super 120B [EU] for EU residency. Ultra is the global OpenRouter path when residency isn't the constraint and you want the bigger model for harder planning loops.

Try it

• "Switch to Nemotron 3 Ultra and build a go-to-market plan — research competitors, model pricing scenarios, and save a memo plus slide outline."
• "Use Nemotron 3 Ultra to refactor this service end-to-end. Keep your plan visible and call tools as you go."
• "Run a swarm with Nemotron 3 Ultra leading — workers gather sources, you synthesize into a board-ready brief."

Heads up

• Workhorse tier, not SOTA — Ultra is ranked and priced as a Workhorse model today. Pick Opus 4.8, Qwen 3.7 Max, or GPT-5.4 when you need the absolute frontier tier.
• Global routing via OpenRouter — the [Global] label means prompts go through OpenRouter's global path. For EU-hosted Nemotron, use Nemotron 3 Super 120B [EU] on Bedrock instead.
• Available on every package — Community through Black and VIP tiers can select it manually or receive it from Auto.
• Tagged "new" until July 4 — the pill in the picker is a freshness flag, nothing functional.