NVIDIA Nemotron 3 Ultra: Accelerating Efficient Reasoning for Long-Running Agents

Overview

NVIDIA has released Nemotron 3 Ultra, a powerful artificial intelligence model designed to handle complex, long-running tasks. This model is built with 55 billion active parameters out of a total of 550 billion. Unlike standard chatbots that answer single questions, Nemotron 3 Ultra is optimized for agentic systems. These are advanced programs that can plan, use tools, read data, delegate tasks to smaller helpers, check their own work, and fix mistakes over many steps. The model combines top-tier reasoning skills with high speed and the ability to adapt to different fields.

Benefits

Nemotron 3 Ultra offers significant improvements in both accuracy and efficiency compared to other models in its class. It achieves a 91% score on the PinchBench, which matches the performance of top-tier models. It also excels in long-term planning, coding, and following instructions. One of its biggest advantages is its ability to handle very long contexts with 95% accuracy, even when processing up to one million tokens. This outperforms many competitors that are limited to much shorter inputs.

Beyond accuracy, the model is engineered for speed and cost savings. It delivers up to five times higher throughput per GPU compared to standard processing on Blackwell hardware. This means it can complete tasks much faster. Additionally, it reduces the cost to complete a task by up to 30% on various benchmarks. This is achieved by using fewer tokens overall and fewer tokens per turn. The model uses a unique hybrid architecture that combines Mamba and Transformer layers. This allows it to handle long sequences efficiently while still remembering specific facts precisely. It also uses a special quantization method called NVFP4, which lets it run on multiple types of NVIDIA GPUs without needing different versions of the model.

Use Cases

Nemotron 3 Ultra is designed for scenarios where an AI needs to work autonomously over a long period. It is ideal for orchestrating complex agent workflows. For example, it can be used to manage software development projects where it needs to write code, run tests, and fix bugs without human intervention. It can also handle knowledge work tasks that require reading large documents and synthesizing information from many sources.

The model is officially supported by agent frameworks like Hermes Agent and OpenClaw. These tools provide the structure needed for multi-turn workflows. Developers can use Nemotron 3 Ultra to build secure, always-on systems that operate safely. It integrates with NVIDIA OpenShell and NemoClaw to ensure that autonomous agents execute tasks without causing errors. The model is also available as a microservice, allowing it to run on various cloud platforms. This flexibility makes it suitable for enterprises that need to deploy AI solutions across different environments.

Pricing

NVIDIA has not disclosed specific pricing details for Nemotron 3 Ultra. The model is released under the OpenMDW-1.1 license, which is a permissive framework designed for open AI model distributions. This license covers the architecture, parameters, documentation, and software. Because the model is fully open, including its weights and training data, developers can adapt it for their own needs. Enterprises can deploy it across various cloud service providers and inference software partners based on their own infrastructure costs.

Vibes

The release of Nemotron 3 Ultra has been well-received for its focus on practical efficiency. The model demonstrates strong performance on benchmarks like SWE-bench and Terminal Bench 2.0, proving its capability in real-world coding and task completion scenarios. The transparency of NVIDIA in releasing training recipes and data has been noted as a positive step for the community. By providing 10 million new supervised fine-tuning samples and 1 million new reinforcement learning tasks, the company has shown a commitment to improving model capabilities through open collaboration. The move to the OpenMDW-1.1 license is seen as a way to reduce licensing ambiguity and encourage broader adoption by developers and enterprises alike.

Additional Information

Alongside Nemotron 3 Ultra, NVIDIA is launching two specialized models to expand its ecosystem. Nemotron 3.5 Content Safety acts as an inference-time guardrail covering 23 safety categories and 12 languages. It can be used for safety testing or to ensure models behave safely. Nemotron 3.5 ASR is a model designed for voice-native agents, enhancing voice input capabilities for more natural interactions.

The training methodology for Nemotron 3 Ultra uses a technique called Multi-Teacher On-Policy Distillation. In this process, the student model learns from over ten specialized teacher models while also generating its own attempts during training. This co-evolution process allows for continuous improvement and progressive specialization across different domains. The model was built on a foundation of 10 trillion tokens, with additional data specifically targeting gaps in legal, general knowledge, and coding domains. This extensive and transparent data pipeline ensures the model is well-rounded and up-to-date with information available through September 30, 2025.