Nemotron 3 Ultra delivers 5x speed for long-running agents

Nemotron 3 Ultra runs long-running agent loops 5x faster and 30% cheaper.

On day zero, NVIDIA Nemotron 3 Ultra is now available on Amazon SageMaker JumpStart, letting teams deploy this open model with a single click. The offering is pitched for frontier reasoning and orchestration in long-running autonomous agents, where planning, tool calls, and self-correction loops can span hundreds of turns. The team reports a 5x speedup in inference for such agentic workloads and up to 30% lower cost, a combination that matters when every turn stacks up tokens and compute.

Nemotron 3 Ultra is a massive yet carefully scoped system. It carries 550 billion total parameters, with 55 billion active parameters per forward pass, and it uses a hybrid Transformer Mamba mixture-of-experts architecture. The model is optimized for the NVFP4 format, a precision setting that the release argues makes the model faster and more cost effective to host. Context length stretches up to 1 million tokens, a scale designed to sustain planning across extended dialogue, multi-step reasoning, and tool calls without collapsing into bottlenecks. In practice, this means agents can sustain planning, tool usage, and self-correction loops across hundreds of turns while maintaining throughput.

The architectural choice at the heart of Nemotron 3 Ultra is the mixture of experts, which activates only 55 billion of the 550 billion parameters for each forward pass. That sparsity is what unlocks the claimed speed and cost benefits even as the model handles very long contexts. In other words, the model can maintain high throughput while engaging in dense, multi-turn reasoning that would strain a dense, fully activated transformer of equivalent quality. By framing inference around select experts, the model remains nimble enough for agentic tasks that require planning, tool delegation, and iterative refinement.

From an engineering standpoint, the JumpStart deployment is a notable constraint breaker. A one-click deployment lowers the barrier to experimentation, letting teams test orchestration-heavy workflows without the upfront burden of managing large model hosting, optimization, and scaling. The platform also anchors the model in an enterprise-grade environment, enabling researchers and product teams to push frontier AI into real product cycles more rapidly than before. The NVFP4 optimization and the long-context capability together position Nemotron 3 Ultra as a potential backbone for autonomous agents that persist across tasks, tools, and evaluation steps rather than delivering a single-shot response.

What practitioners should watch next goes beyond raw speed and price. First, the MoE routing and expert load balancing, even with a capped active parameter set, will be a critical reliability signal as teams scale to longer-running tasks. Second, the million-token context length tests memory management and streaming of results; developers should design robust chunking and tool-calling strategies to avoid latency spikes. Third, JumpStart deployment invites rapid experimentation, but enterprises will want clear cost accounting and governance around model selection, rate limits, and monitoring. Finally, as an open model designed for frontier use, ecosystem compatibility and tooling around agent orchestration will determine how quickly teams can translate these gains into real products.

In short, Nemotron 3 Ultra crystallizes a core engineering lesson: sparing the active parameter count while preserving context length and orchestration capability can unlock practical gains in speed and cost for the kind of multi-turn, tool-enabled agents that enterprise teams are increasingly building.