NVIDIA Nemotron 3 Ultra: The Fastest Open Model for Agent AI

On June 4, 2026, NVIDIA released the Nemotron 3 Ultra, the flagship model of the Nemotron 3 family and by far one of the most productive open language models in the world. This is not just another big LLM: the architecture, the principles of learning and the focus of application are fundamentally different.

---

## Why do you even need to

Single chatbots are a thing of the past. Modern AI is a long chain of agents: the scheduler calls tools, tools return data, data is transmitted further, subtasks are delegated to sub-agents. With each step, the context window grows, the cost of inferencing increases, and the risk of loss of purpose accumulates.

Nemotron 3 Ultra is designed for this scenario: complex orchestration, long context, high speed – with accuracy at the level of the best open models.

---

## What's Inside: Architecture

### 550 billion parameters, 55 active parameters

The Nemotron 3 Ultra is a Mixture-of-Experts model. The total parameters are 550B, but only 55B is activated for each token. This means: the power of a large model with significantly lower computational costs for inference.

### Mamba + Transformer hybrid

Instead of a pure Transformer architecture, NVIDIA uses a hybrid: most layers are Mamba (long sequence processing with sub-square complexity), several layers of Attention are left to accurately extract information from context. This provides a context window of up to 1 million tokens at a reasonable cost.

### LatentMoE

Own routing mechanism of experts from NVIDIA. Compared to the standard MoE, it improves the accuracy of the model due to improved distribution of tokens among specialized experts.

### MTP (Multi-Token Prediction)

The model predicts multiple tokens simultaneously, which works as built-in speculative decoding. The adoption of the first two predicted tokens is around 97% – this significantly speeds up generation, especially in long sessions.

### NVFP4 quantization

The model is pretrained in NVFP4 format - 4-bit floating point accuracy from NVIDIA. This provides up to 5x higher throughput compared to similar models on the GB200/GB300 GPU architecture with minimal quality losses (deviation from BF16 is less than 0.4%).

---

## How to train

### Pre-education

The model is trained on 16+ trillion tokens, including:

• 173 billion GitHub code tokens (data through September 2025)
• synthetic datasets for legal purposes
• datasets to improve actual accuracy and complex reasoning scenarios

### Post-training: SFT → RL → MOPD

After pre-training, the model underwent a three-stage pipeline training:

1. SFT - supervised fine tuning for basic alignment
2. RL – Multi-environmental reinforcement learning (15 new RL environments in this release, 55 in total)
3. MOPD (Multi-teacher On-Policy Distillation) - knowledge distillation from 10+ specialized models-teachers. Each round: the student rollouts, the teachers give tight feedback, the knowledge merges back. NVIDIA conducted 2 iterations of the MOPD for Nemotron 3 Ultra.

---

## Performance: Accuracy and Speed

### Accuracy on key benchmarks

Бенчмарк	Nemotron 3 Ultra	GLM 5.1 (744B)	Kimi K2.6 (1T)	Qwen 3.5 (397B)
Agent Productivity (PinchBench)	91%	84%	91%	89%
Instruction Following (IFBench)	82%	77%	74%	78%
Long Context (Ruler @1M)	95%	N/A (макс. 256K)	N/A (макс. 256K)	90%
Professional Work (ProfBench)	56%	46%	56%	53%
Coding (Terminal-Bench 2.0)	54%	64%	67%	53%
SWE-Bench Verified	71.9	—	—	—
IOI 2025 (конкурсное программирование)	570.0	—	—	—

According to NVIDIA, the result of 570.0 at IOI 2025 corresponds to the top 3 level among people in competitive programming.

### Inferencing speed

This is the main advantage of the model. On the 8K entry/64K exit token setting, Nemotron 3 Ultra outperforms competitors:

• in 5.9x faster than GLM-5.1-754B
• in 4.8x faster than Kimi-K2.6-1T
• in 1.6x faster than Qwen-3.5-397B

It’s not just statistics – for agent systems, where a single session can generate hundreds of thousands of tokens, speed directly affects cost and applicability.

---

## Three modes of reasoning

The model supports reasoning budget management in runtime:

• Reasoning off - direct answer without a chain of reasoning, for simple tasks
• Regular - Standard reasoning
• Medium - Extended reasoning for complex problems

This allows you to balance between speed and depth of analysis during the operation of the agent.

---

## Openness: What is available

NVIDIA releases Nemotron 3 Ultra completely openly under the OpenMDW-1.1 license:

• model weight (BF16 and NVFP4)
• 10+ trillion training data tokens
• instructional
• 50 million SFT samples (10M new in this release)
• 2 million RL tasks (1M new)
• 55 RL environments (15 new)

It's rare for a model of this scale to open either weights or a piece of data.

---

## What tasks are suitable for

The Nemotron 3 Ultra was designed for scenarios where other models are either too expensive or too slow

• *Coding ** - Supporting architectural solutions through long sessions, debugging complex systems
• **Research ** - synthesis of information from hundreds of sources with full context
• Autonomous agents – orchestration of multi-way workflows with instrument call
• *Verification – Testing complex systems (such as chip design) against thousands of constraints

---

## Where to start

The model is available through:

• Hugging Face - weights for self-deployment
• NVIDIA NIM - ready-made inference via API
• build.nvidia.com - NVIDIA cloud platform

For maximum performance, the GB200 GPU architecture with TRT-LLM is recommended.

---

## Outcome

Nemotron 3 Ultra is NVIDIA’s answer to the practical problem of Agent AI: how to get the quality of a frontier model at cost and production speed. Hybrid Mamba architecture, NVFP4-quantization and MTP-acceleration give up to 5-6x advantages throughput at comparable accuracy with models that weigh 2-3 times more.

Complete openness – weights, data, recipes – makes Nemotron 3 Ultra a rare case where the research community gets not just a model, but the entire stack for playback and learning.