Building serious agentic AI means solving a problem most teams hit eventually: not every task in your workflow requires the same model. Some decisions need deep reasoning. Most don’t. The teams that figure out how to match model capability to task complexity are the ones that build systems that actually work in production: efficiently, reliably, and at scale.

That’s the problem NVIDIA Nemotron 3 Ultra was built for. And starting today, it’s available to run, thanks to Anaconda’s acquisition of Outerbounds.

A model built for the hard part

Nemotron 3 Ultra is NVIDIA’s largest open frontier reasoning model: 550B parameters with 55B active per token, built on a hybrid Mamba-Transformer Mixture of Experts (MoE) architecture with a 1M token context window.

Designed for long-running agentic workflows, Nemotron 3 Ultra delivers up to 5x faster inference and up to 30% lower cost for agentic tasks compared to other open models in its class.

The architecture matters: rather than activating the full model for every request, it routes each token through the experts it actually needs. You get frontier-level reasoning without the compute overhead of a dense model at this scale.

NVIDIA built Ultra specifically for the orchestration layer of agentic systems: long-running agent workflows, long-horizon planning, complex multi-step reasoning, synthesis across large volumes of information. These are tasks where getting it wrong cascades through everything downstream.

The Nemotron family as a system

If you’ve already built with Nemotron 3 Nano on Anaconda, you know what it’s good for: fast, efficient sub-agents handling high-volume tool calls, retrieval, and validation at low cost.

Ultra completes the picture. This is what compound AI systems look like in practice: multiple specialized models working together within a single project, each handling the part of the workflow it’s best suited for. Outerbounds is built for exactly this: run Ultra and Nano together inside a single project structure, with shared context, unified observability, and no infrastructure stitching required.

  • Ultra sits at the orchestration layer: planning, reasoning, synthesis, decisions that require genuine depth
  • Nano handles the volume: tool calls, API actions, validation, anything that needs speed and efficiency at scale

This is the missing piece that makes the Nemotron family a full agentic stack. Having both available on the same platform means you build that architecture without routing requests across different providers, managing separate governance policies, or stitching together infrastructure that wasn’t designed to work together.

Why this matters for enterprise AI

What makes this combination genuinely interesting for enterprises is what is possible when running AI models inside a governed platform.

From Ultra for frontier reasoning and orchestration, to Super for general-purpose agent workflows, Nano for high-volume execution, and Nano Omni for multimodal understanding, organizations can match model capability to task complexity while maintaining a unified operational environment.

All Nemotron models are released with open weights, recipes, and training assets, enabling organizations to customize, fine-tune, and deploy models within their own environments while maintaining control over their data and AI infrastructure.

With Outerbounds handling orchestration and the Anaconda Platform handling governance, the goal is straightforward: a workflow that uses Ultra for planning and Nano for execution should inherit the same security controls, audit trails, and compliance policies your organization has already configured. Your data stays in your environment.

That’s a meaningfully different proposition from accessing these models through a public API. You get frontier capability without giving up control.

What's next

Running Ultra on the Anaconda Platform is the first step. Coming soon: full model catalog integration with AIBOM documentation, benchmark data at each quantization level, and a governed one-line deployment experience.

The end state: the same trust and transparency Anaconda brings to every model in the catalog, applied to the most capable open reasoning model available today.

Nemotron 3 Ultra is available on Outerbounds now. Contact us to learn more →