Nvidia drops Nemotron 3 Ultra to supercharge long-running agents

Nemotron 3 Ultra is an open-source Mixture-of-Experts model from Nvidia built specifically for complex orchestration and long-running autonomous agents. It packs 550B total parameters while activating just 55B parameters per forward pass, paired with an expansive 1M token context window.

In multi-turn agent workflows like long coding sessions or multi-step research, standard models often get bogged down by exploding context histories, driving up compute bills and drifting away from goals. Nemotron 3 Ultra targets this by cutting tokens used per turn, slashing agent task costs by up to 30% while pushing 5x higher throughput than comparable open models.

How it works and key benchmarks

The model drops the traditional pure Transformer layout in favor of a hybrid Mamba-Transformer architecture. Mamba layers handle sequence throughput across long context stretches, while interleaved Transformer layers ensure precise recall when an agent needs specific facts buried deep in memory.

It is optimized natively for NVFP4 precision, allowing the exact same checkpoint to deploy across Nvidia Ampere, Hopper, and Blackwell GPUs. Along with multi-token prediction, this format delivers massive speedups, achieving up to 5x higher throughput per GPU on Blackwell compared to standard BF16.

On the 1M token Ruler benchmark, Nemotron 3 Ultra scored 95%, beating Qwen3.5 at 90%, while models like GLM 5.1 and Kimi K2.6 topped out at 256K. On PinchBench it reached 91%, and on SWE-bench Verified it maintained scores between 65% and 70.4% across multiple harnesses including OpenHands, Pi, and Hermes Agent.

Training leveraged Multi-Teacher On-Policy Distillation, where over ten domain-specialist teacher models scored student rollouts iteratively. Nvidia layered 212B new tokens on top of its 10T pre-training base, including 4B synthetic legal tokens, 35B synthetic wiki tokens, and 173B GitHub tokens refreshed through late 2025.

Alongside the flagship model, Nvidia also launched Nemotron 3.5 Content Safety, an open 4B guardrail covering 23 categories and 12 languages, and Nemotron 3.5 ASR, a sub-100ms streaming speech model supporting over 40 languages.

Hardware specs and how to run it

Nemotron 3 Ultra is released under the Linux Foundation's OpenMDW-1.1 license, covering weights, NeMo training recipes, and millions of curated SFT and RL samples.

With 550B total parameters, self-hosting requires significant enterprise hardware. On the cloud side, Amazon SageMaker JumpStart added zero-day one-click deployment across multi-GPU instances like ml.p5en.48xlarge, ml.p5.48xlarge, and ml.g7e.48xlarge.

It is supported natively by serving engines including vLLM, SGLang, and TRT-LLM, as well as Nvidia NIM containers. Developers looking to experiment can try it directly on build.nvidia.com, through OpenRouter, or via Perplexity Pro.

For execution, Nvidia provides NemoClaw and the OpenShell runtime preview, enabling developers to run autonomous agents in a secure sandbox alongside frameworks like Hermes Agent and OpenClaw.

Why it matters

If you are building autonomous agents that need to plan across hundreds of turns, write complex code, or call nested tools, Nemotron 3 Ultra is well worth testing. Activating only 55B parameters keeps inference snappy, while the hybrid design and 1M context keep multi-turn reasoning on track without breaking the budget.

The clear bottleneck is hardware footprint. Even with 4-bit quantization, running a 550B parameter beast locally is out of reach for individual developers, making hosted inference APIs and managed cloud endpoints the realistic path forward.

For the broader ecosystem, Nvidia is doing more than releasing a raw model checkpoint. By open-sourcing the distillation pipelines, safety guardrails, and secure execution runtime together, it shifts the open-source agent competition from single-turn benchmark racing to full-stack, cost-effective runtime execution.