NVIDIA Nemotron 3 Ultra 550B Overview

Model Architecture and Training

NVIDIA's Nemotron 3 Ultra is a 550 billion parameter mixture-of-experts (MoE) model with 55 billion active parameters. It is designed specifically for agentic workflows, including tool calling, coding, and long-horizon reasoning. The model supports a 1-million token context window and multi-token prediction.

NVIDIA utilized multi-teacher distillation to achieve high performance across diverse tasks. They trained specialized teacher models for code, tool use, and instruction following, then distilled these capabilities into the final model. This approach reportedly yields superior results compared to training a single model on a combined dataset. Additionally, the model underwent post-training on agent trajectories derived from harnesses like Open Claw and Hermes to improve task completion and error recovery.

Reasoning and Tool Use

The model exposes reasoning capabilities through an OpenAI-compatible API, allowing users to toggle chain-of-thought processing via an enable_thinking flag. Users can control the reasoning depth using a reasoning_budget parameter, which limits the number of thinking tokens generated. In practice, the model remains succinct even with high budgets, prioritizing speed and task completion. It demonstrates strong capabilities in multi-step tool calling, effectively processing tool outputs to determine subsequent actions in an agent loop.

Performance and Benchmarks

In agent-focused benchmarks like Pinchbench, Nemotron 3 Ultra performs competitively against proprietary models like Claude 3.5 Opus. It notably outperforms larger models, such as the 1-trillion parameter GLM 5.1, while maintaining higher throughput, reaching speeds over 300 tokens per second. NVIDIA has committed to releasing datasets and reinforcement learning (RL) environments used in the model's training, providing a transparent recipe for organizations to fine-tune custom versions of the model for specific enterprise use cases.