ai/nemotron-3-ultra repository overview

⁠Read our How to Run Nemotron 3 Ultra Guide!⁠

See Unsloth Dynamic 2.0 GGUFs⁠ for our quantization benchmarks.

⁠ ⁠ ⁠

You can view KLD Divergence Benchmarks, in our guide⁠.

⁠NVIDIA-Nemotron-3-Ultra-550B-A55B

⁠ ⁠ ⁠ ⁠

⁠ ⁠

⁠

⁠Model Summary

⁠Quick Start

For more details on how to deploy and use the model - see the Quick Start Guide⁠ below!

⁠Model Overview

Model Developer: NVIDIA Corporation

Model Dates: December 2025 - April 2026

Data Freshness:

• The post-training data has a cutoff date of May 2026.
• The pre-training data has a cutoff date of September 2025.

⁠What is Nemotron?

NVIDIA Nemotron™ is a family of open models with open weights, training data, and recipes, delivering leading efficiency and accuracy for building specialized AI agents.

⁠Description

Nemotron-3-Ultra-550B-A55B-BF16 is a frontier-scale large language model (LLM) trained by NVIDIA, designed to deliver strong agentic, reasoning, and conversational capabilities. It is optimized for the most demanding workloads, including complex multi-step agents, long-context analysis, and high-accuracy reasoning over code, math, and science. Like other models in the family, it responds to user queries and tasks by first generating a reasoning trace and then concluding with a final response. The model's reasoning capabilities can be configured through a flag in the chat template.

The model employs a hybrid Latent Mixture-of-Experts (LatentMoE) architecture, utilizing interleaved Mamba-2 and MoE layers, along with select Attention layers. Like the Super model, the Ultra model incorporates Multi-Token Prediction (MTP) layers for faster text generation and improved quality, and it is trained using an NVFP4 pre-training recipe to maximize compute efficiency. The model has 55B active parameters and 550B parameters in total.

The supported languages include: English, French, Spanish, Italian, German, Japanese, Korean, Hindi, Korean, Brazilian Portuguese, and Chinese.

This model is ready for commercial and non-commercial use.

⁠License/Terms of Use

Governing Download Terms: Use of this model is governed by the OpenMDW License Agreement, version 1.1⁠ (OpenMDW-1.1).

⁠Benchmarks

All evaluation results were collected via Nemo Evaluator SDK⁠. We used three main evaluation harnesses: Nemo Gym⁠, Nemo Skills⁠, and Harbor⁠ with extended sandboxing support via AWS ECS on Nemo Evaluator. In addition, the evaluations also used dedicated open-source packaged containers for ScaleAI Multi Challenge Multi Turn Instruction Following and KernelBench. For reproducibility purposes, more details on the evaluation settings and pinned containers can be found in the Nemo Evaluator SDK examples folder⁠ and the reproducibility tutorial for Nemotron 3 Ultra⁠.

The following benchmarks are not onboarded yet in our open source tools and for these we used either their official open source implementation or otherwise an internal scaffolding that we plan to open source in the future: BrowseComp with Search, Tau Bench 3, ProfBench with Search, PinchBench, Vals.ai, LongBench v2.

⁠Deployment Geography: Global

⁠Use Case

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is a frontier-scale general purpose reasoning and chat model intended to be used in English, Code, and supported multilingual contexts. This model is optimized for complex agentic workflows, long-context reasoning, and high-stakes analytical workloads. It is intended to be used by developers designing AI Agent systems, chatbots, RAG systems, and other AI-powered applications. This model is also suitable for complex instruction-following tasks and long-context reasoning over very large documents and codebases.

⁠Release Date

Hugging Face - 06/04/2026 via Hugging Face⁠

⁠Reference(s)

• NVIDIA Nemotron 3 model family on Hugging Face⁠
• NVIDIA Nemotron 3 Ultra Technical Report⁠

⁠Model Architecture

• Architecture Type: Mamba2-Transformer Hybrid Latent Mixture of Experts (LatentMoE) with Multi-Token Prediction (MTP)
• Network Architecture: Nemotron Hybrid LatentMoE
• Number of model parameters: 550B Total / 55B Active

⁠Model Design

The model utilizes the LatentMoE architecture, where tokens are projected into a smaller latent dimension for expert routing and computation, improving accuracy per byte. The Ultra model is pre-trained using an NVFP4 recipe — sharing the quantization-aware pre-training approach pioneered in the Nemotron 3 family. The majority of linear layers use NVFP4 for weights, activations, and gradients, while select layers (including latent projections, MTP layers, QKV/attention projections, and embeddings) are maintained in BF16 or MXFP8 for training stability. The model includes Multi-Token Prediction (MTP) layers using a shared-weight design across prediction heads. This improves training signal quality, enables faster inference via native speculative decoding, and supports more stable autoregressive drafting at longer draft lengths compared to independently trained offset heads.

⁠Training Methodology

Stage 1: Pre-Training

• NVIDIA-Nemotron-3-Ultra-550B-A55B-Base-BF16⁠ model was pre-trained for approximately 20T tokens using crawled and synthetic code, math, science, and general knowledge data. Training leveraged an NVFP4 recipe for efficiency. All datasets are disclosed in the Training and Evaluation Datasets⁠ section of this document. Major portions of the pre-training corpus are released in the Nemotron-Pre-Training-Datasets⁠ collection.
• Software used for pre-training: Megatron-LM⁠

Stage 2: Supervised Fine-Tuning

• The model was further fine-tuned on synthetic code, math, science, tool calling, instruction following, structured outputs, and general knowledge data. This stage incorporated data designed to support long-range retrieval and multi-document aggregation. All datasets are disclosed in the Training and Evaluation Datasets⁠ section of this document. Major portions of the fine-tuning corpus are released in the Nemotron-Post-Training-v3⁠ collection. Data Designer⁠ is one of the libraries used to prepare these corpora.

Stage 3: Reinforcement Learning

• The model underwent multi-environment reinforcement learning using asynchronous GRPO (Group Relative Policy Optimization) across math, code, science, instruction following, multi-step tool use, multi-turn conversations, and structured output environments. It utilized an asynchronous RL architecture that fully decouples training from inference across separate GPU devices, leveraging in-flight weight updates and MTP to accelerate rollout generation. Conversational quality was further refined through RLHF. All datasets are disclosed in the Training and Evaluation Datasets⁠ section of this document. The RL environments and datasets are released as part of NeMo Gym⁠.
• Software used for reinforcement learning: NeMo RL⁠, NeMo Gym⁠

Stage 4: Multi-Domain On-Policy Distillation (MOPD)

• The model underwent Multi-Domain On-Policy Distillation (MOPD) to improve reasoning across many task types while staying efficient. This technique uses strong teacher models to guide training on the model's own generated attempts (on-policy rollouts), helping recover accuracy and improve performance across coding, math, instruction following, tool use, and agentic workflows. By distilling teacher signal onto the student's own trajectories rather than offline traces, MOPD better aligns the student's behavior with what it would actually produce at inference time, yielding stronger gains than purely off-policy distillation.

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 model is a result of the above work.

The end-to-end training recipe is available in the NVIDIA Nemotron Developer Repository⁠. Evaluation results can be replicated using the NeMo Evaluator SDK⁠. Data Designer⁠ is one of the libraries used to prepare the pre and post training datasets. More details on the datasets and synthetic data generation methods can be found in the technical report NVIDIA Nemotron 3 Ultra Technical Report⁠.

⁠Input

• Input Type(s): Text
• Input Format(s): String
• Input Parameters: One-Dimensional (1D): Sequences
• Other Properties Related to Input: Maximum context length up to 1M tokens. Supported languages include: English, French, Spanish, Italian, German, Japanese, Korean, Hindi, Korean, Brazilian Portuguese, and Chinese

⁠Output

• Output Type(s): Text
• Output Format: String
• Output Parameters: One-Dimensional (1D): Sequences
• Other Properties Related to Output: Maximum context length up to 1M tokens

Our AI models are designed and optimized to run on NVIDIA GPU-accelerated systems. By leveraging NVIDIA's hardware (e.g. GPU cores) and software frameworks (e.g., CUDA libraries), the model achieves faster training and inference times compared to CPU-only solutions.

⁠Software Integration

• Runtime Engine(s): NeMo 26.04.01
• Supported Hardware Microarchitecture Compatibility: NVIDIA Ampere - A100; NVIDIA Blackwell; NVIDIA Hopper - H100-80GB
• Operating System(s): Linux

The integration of foundation and fine-tuned models into AI systems requires additional testing using use-case-specific data to ensure safe and effective deployment. Following the V-model methodology, iterative testing and validation at both unit and system levels are essential to mitigate risks, meet technical and functional requirements, and ensure compliance with safety and ethical standards before deployment.

⁠Model Version(s)

• v1.0 - GA

The Ultra BF16 checkpoint is a frontier-scale model. The minimum recommended hardware is:

• Single-node: 8× B200 (≈1.5 TB aggregate HBM — fits BF16 weights plus KV cache with headroom)
• Multi-node: ≥8 GPUs across H100 / H200 / GB200 / GB300, orchestrated with Ray v2

All deployment snippets below default to port 8000, with chunked prefill and MTP (5 speculative tokens) enabled.

⁠Multi-Node Setup with Ray (recommended for multi-node BF16)

The recommended multi-processing backend for multi-node BF16 deployments is Ray v2. Below is a template for launching a Ray cluster:

# Set the IP for the head node in RAY_HEAD_IP
export RAY_HEAD_IP=<head_node_ip>
export RAY_PORT=6379
export RAY_ADDRESS=${RAY_HEAD_IP}:${RAY_PORT}

# Start Ray head node (vLLM/SGLang will run on this node)
ray start --head --node-ip-address=${RAY_HEAD_IP} --port=${RAY_PORT}

# Start Ray worker node(s)
ray start --address=${RAY_HEAD_IP}:${RAY_PORT} --block

# Verify Ray cluster is ready
ray status --address=${RAY_HEAD_IP}:${RAY_PORT}

ray[cgraph] is required: uv pip install "ray[cgraph]"

⁠vLLM

Recommended container: vllm/vllm-openai:v0.22.0.

For more detailed information, please see this cookbook⁠.

export MODEL_CKPT=PATH/TO/MODEL/CHECKPOINT

8× B200 single-node deployment:

docker run -d --name nemotron-ultra-vllm \
  --gpus all \
  --ipc=host \
  --network=host \
  --shm-size=16g \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  -v $MODEL_CKPT:/model:ro \
  -e VLLM_WORKER_MULTIPROC_METHOD=spawn \
  -e SAFETENSORS_FAST_GPU=1 \
  -e NVIDIA_TF32_OVERRIDE=1 \
  -e VLLM_LOGGING_LEVEL=INFO \
  vllm/vllm-openai:v0.22.0 \
  /model \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name nvidia/nemotron-3-ultra \
  --trust-remote-code \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --dtype bfloat16 \
  --max-model-len 262144 \
  --gpu-memory-utilization 0.90 \
  --max-num-seqs 16 \
  --max-num-batched-tokens 32768 \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --reasoning-parser nemotron_3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --mamba-ssm-cache-dtype float16 \   
  --mamba-backend flashinfer \     
  --enable-mamba-cache-stochastic-rounding \   
  --mamba-cache-philox-rounds 5 \
  --speculative-config '{"method": "nemotron_h_mtp", "num_speculative_tokens": 5}' \
  --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 96}'

Multi-node deployment (e.g. 2× 4×GB300 with Ray):

After launching the Ray head and worker per the multi-node setup above:

# Run on Ray head node
vllm serve $MODEL_CKPT \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name nvidia/nemotron-3-ultra \
  --tensor-parallel-size 8 \
  --distributed-executor-backend ray \
  --trust-remote-code \
  --dtype bfloat16 \
  --gpu-memory-utilization 0.90 \
  --max-model-len 262144 \
  --max-num-seqs 256 \
  --max-num-batched-tokens 32768 \
  --enable-chunked-prefill \
  --enable-prefix-caching \
  --reasoning-parser nemotron_3 \
  --mamba-ssm-cache-dtype float16 \   
  --mamba-backend flashinfer \     
  --enable-mamba-cache-stochastic-rounding \   
  --mamba-cache-philox-rounds 5 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --speculative-config '{"method": "nemotron_h_mtp", "num_speculative_tokens": 5}' \
  --kv-cache-dtype fp8 \
  --model-loader-extra-config '{"enable_multithread_load": true, "num_threads": 96}' \
  --compilation-config '{"pass_config": {"fuse_allreduce_rms": false}}' \
  --distributed-timeout-seconds 3600

Context length defaults to 256k above. To use up to 1M, set VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 and --max-model-len 1048576.

Useful environment variables: VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm, VLLM_FLASHINFER_MOE_BACKEND=latency (TRTLLM-Gen) or VLLM_FLASHINFER_MOE_BACKEND=throughput (CUTLASS).

⁠SGLang

Container (tested on 8× B200):

docker pull lmsysorg/sglang:v0.5.11

For more detailed information, please see this cookbook⁠.

8× B200 single-node deployment (BF16, chunked prefill + MTP on by default):

docker run -d --name nemotron-ultra-sglang \
  --gpus all \
  --cap-add SYS_NICE \
  --ipc=host \
  --network=host \
  --shm-size=16g \
  --ulimit memlock=-1 \
  --ulimit stack=67108864 \
  -v $MODEL_CKPT:/model:ro \
  -e SAFETENSORS_FAST_GPU=1 \
  -e NVIDIA_TF32_OVERRIDE=1 \
  -e SGLANG_DISABLE_DEEP_GEMM=1 \
  lmsysorg/sglang:v0.5.11 \
  python3 -m sglang.launch_server \
  --model-path /model \
  --host 0.0.0.0 \
  --port 8000 \
  --served-model-name nvidia/nemotron-3-ultra \
  --tp-size 8 \
  --ep-size 8 \
  --context-length 262144 \
  --mem-fraction-static 0.85 \
  --chunked-prefill-size 32768 \
  --fp8-gemm-backend triton \
  --moe-runner-backend triton \
  --mamba-scheduler-strategy no_buffer \
  --disable-piecewise-cuda-graph \
  --reasoning-parser nemotron_3 \
  --tool-call-parser qwen3_coder \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 5 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 5 \
  --trust-remote-code \
  --log-level info

Context length defaults to 256k above. To use up to 1M, set SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 and --context-length 1048576.

Tool calls + reasoning parsing: when calling the chat completions endpoint with tools, you must set "chat_template_kwargs": {"enable_thinking": true, "force_nonempty_content": true} in the request body to parse both reasoning and tool calls correctly.

⁠**Tens

…(truncated — see the full README on HuggingFace)