NVIDIA Nemotron 3 Ultra, now live on Lightning AI
Introducing NVIDIA Nemotron 3 Ultra
Developers are building agents that plan, reason, delegate work, check results, and complete tasks autonomously. These agents operate continuously across multi-step workflows, requiring deep reasoning, real-time planning, and adaptive decision-making at scale.
But long-running agentic workflows introduce a new kind of complexity.
Most multi-agent systems today rely on expensive proprietary frontier models for orchestration and reasoning. As these systems scale, three issues consistently show up:
Reasoning bottlenecks
Frontier models are expensive per inference, becoming prohibitive at scale across long chains of reasoning
Context window constraints
Orchestration tasks require long-context understanding, and proprietary models come with licensing friction and operational lock-in
Cost inefficiency
A single reasoning model doesn't fit all parts of the agent workflow, yet enterprises end up paying frontier model pricing for every task
Today, we're excited to support Nemotron 3 Ultra on Lightning, giving developers an open, frontier-class reasoning model built for orchestration in long-running agent systems.
Enter NVIDIA Nemotron 3 Ultra
NVIDIA Nemotron 3 Ultra is a frontier-class reasoning and orchestration model that serves as the "brain" of an agent system.
With 550B parameters (55B active via Mixture of Experts (MoE)) and up to 1M context length, it maintains complex reasoning across extended workflows, enabling agents to plan deeply, call tools with precision, validate results, and adapt in real time.
Instead of relying solely on expensive proprietary models or underpowered open alternatives, developers can deploy Nemotron 3 Ultra as their reasoning backbone, paired with specialized models for execution, perception, and safety.
Reasoning without the frontier model tax
Long-running agent systems are expensive by default.
Every reasoning step requires either expensive proprietary model calls or cascading errors from under-capable alternatives.
NVIDIA Nemotron 3 Ultra replaces the need for multiple proprietary calls with a single open, efficient reasoning model, reducing operational costs while enabling deeper, longer reasoning chains.
Using this model in agent orchestration, teams can achieve:
- Significant reduction in per-inference reasoning costs compared to proprietary frontier models
- Extended reasoning chains without hitting context window limits (up to 1M tokens)
- Faster task completion through improved planning and validation cycles
- Full control over model updates and custom fine-tuning for domain-specific workflows
Because of its efficient MoE architecture, Nemotron 3 Ultra delivers 5x higher throughput compared to other open models in its class and up to 30% lower cost for agentic tasks, enabling enterprises to run reasoning-intensive workflows at production scale.
Open, frontier-class, and production-ready
Nemotron 3 Ultra is released with open weights, data, and training recipes, enabling full transparency and control.
Teams can deploy it across cloud, on-prem, or hybrid environments, maintain full control over sensitive reasoning data, and customize the model for domain-specific agent orchestration. This aligns with a growing shift toward systems of models, where specialized reasoning, execution, perception, and safety models work together efficiently in production environments.
Run NVIDIA Nemotron 3 Ultra on Lightning
Building production-ready long-running agents requires more than just a model.
Lightning provides the unified orchestration platform needed to deploy Nemotron 3 Ultra and coordinate it with specialized models across your entire agent system. With Lightning, developers can:
- Deploy multi-model agent stacksIntegrate Ultra for reasoning and orchestration alongside Nemotron 3 Nano for execution, Nemotron 3.5 ASR for voice, and safety models for guardrails, all from one platform
- Scale long-context reasoningHandle Nemotron 3 Ultra's 1M context length with production-grade inference optimization and monitoring
- Deploy anywhereRun agents on-prem, in hybrid environments, or across multiple clouds with seamless switching and no vendor lock-in
- Optimize for your use caseAccess Lightning's optimization team for custom tuning of Ultra's MoE architecture to your specific workflows
- Maintain enterprise controlComprehensive security, compliance, governance, and audit logging for regulated deployments
Each model handles its specialized function, allowing teams to build efficient multi-agent systems while maintaining full transparency and control over sensitive reasoning data.
The next phase of agent systems
Agent systems are becoming longer-running, more autonomous, and deeply integrated into enterprise workflows.
NVIDIA Nemotron 3 Ultra introduces a new layer in the reasoning stack: an open frontier-class model for orchestration and planning that keeps agents capable of complex decision-making without proprietary model dependency.
Combined with platforms like Lightning, developers can build systems that are faster, more cost-efficient, and easier to operate in production.
Get started
Build long-running agents with frontier reasoning, orchestrate complex multi-step workflows, and deploy scalable agentic AI systems - all in one platform.
Try NVIDIA Nemotron 3 Ultra on Lightning and start building the next generation of autonomous agent systems.