NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16#

Chat Paper Pre-Training Datasets Post-Training Datasets

Homepage Discord

License

Image: image

Model Summary#

Quick Start#

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

For running Nemotron 3 Ultra on a smaller footprint, please see: NVIDIA-Nemotron-3-Ultra-550B-A55B-NVFP4

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, 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#

| Benchmark | N-3-Ultra 550B-A55B | MiniMax-2.7 230B-A10B | GLM-5.1 744B-A40B | Kimi-K2.6 1T-A32B | Qwen-3.5 397B-17B | DS-v4-Pro 1.6T-A49B | DS-v4-Flash 284B-A13B |

| :--- | :---: | :---: | :---: | :---: | :---: | :---: | :---: |

| Agentic | | | | | | | |

| Terminal Bench 2.1 | 56.4 | 55.5 | 59.3 | 67.2 | 49.9 | 49.2 | 54.2 |

| GDPVal | 46.7 | 47.6 | 54.7 | 50.4 | 34.6 | 54.6 | 50.2 |

| SWE-Bench Verified | 70.7 | 75.3 | 76.2 | 75.7 | 73.6 | 74.5 | 73.5 |

| SWE-Bench Multilingual | 67.7 | 71.8 | 74.8 | 77.1 | 70.9 | 76.5 | 75.0 |

| ProfBench (Search) | 56.0 | 52.0 | 46.0 | 56.0 | 53.0 | 59.9 | 57.0 |

| PinchBench | 90.0 | 77.6 | 81.2 | 90.2 | 86.6 | 88.6 | 91.3 |

| TauBench V3 | | | | | | | |

|   Airline | 81.5 | 75.3 | 85.0 | 85.8 | 76.5 | 80.8 | 80.8 |

|   Retail | 86.4 | 84.9 | 84.1 | 82.9 | 88.5 | 88.9 | 89.1 |

|   Telecom | 92.9 | 89.6 | 96.9 | 97.8 | 98.0 | 96.3 | 98.3 |

|   Banking | 22.6 | 14.6 | 12.8 | 23.1 | 20.9 | 25.9 | 26.7 |

|   Average | 70.9 | 66.1 | 69.7 | 72.4 | 71.0 | 73.2 | 73.7 |

| BrowseComp | 44.4 | 54.1 | 59.4 | 61.3 | 40.5 | 59.4 | 46.9 |

| Vals.ai Financial Agent 1.1 | | | | | | | |

|   without web search | 60.1 | 51.3 | 60.2 | 54.0 | 61.3 | 58.9 | 58.4 |

|   with web search | 53.7 | 50.5 | 60.7 | 58.8 | 59.0 | 62.3 | 60.1 |

| Reasoning and Knowledge | | | | | | | |

| IOI 2025 | 570.0 | -- | 456.5 | 585.0 | 441.3 | 580.1 | -- |

| LiveCodeBench (v6) | 89.0 | 77.2 | 85.7 | 90.2 | 79.3 | 92.5 | 90.9 |

| IMOAnswerBench (no tools) | 88.6 | 68.3 | 86.8 | 91.1 | 83.1 | 93.0 | 91.1 |

| IMOAnswerBench (with tools) | 92.3 | 75.1 | 91.1 | 93.71 | 84.51 | 85.4 | 89.6 |

| Apex-Shortlist (no tools) | 74.9 | 28.9 | 71.1 | 77.4 | 61.4 | 85.8 | 82.4 |

| Apex-Shortlist (with tools) | 84.8 | 51.9 | 79.0 | 73.2 | 60.4 | 86.5 | 82.0 |

| GPQA (no tools) | 87.0 | 86.6 | 86.1 | 91.0 | 87.1 | 87.8 | 88.5 |

| SciCode (subtask) | 44.6 | 38.3 | 47.7 | 52.0 | 48.0 | 50.5 | 48.2 |

| HLE (no tools) | 26.7 | 23.1 | 27.2 | 34.8 | 28.5 | 37.7 | 32.2 |

| HLE (with tools) | 37.4 | -- | 50.4 | 54.0 | 48.3 | 48.2 | 45.1 |

| CritPt (no tools) | 3.1 | 0.6 | 3.7 | 9.1 | 2.4 | 14.0 | 10.6 |

| MMLU-Pro | 86.8 | 81.9 | 85.9 | 88.1 | 88.3 | 87.5 | 86.4 |

| OmniScience Accuracy | 24.1 | 20.5 | 31.3 | 35.5 | 35.9 | 46.8 | 39.9 |

| OmniScience Non-Hallucination | 78.7 | 74.4 | 66.8 | 67.1 | 7.4 | 5.7 | 2.8 |

| Chat & Instruction Following | | | | | | | |

| IFBench (prompt loose) | 81.7 | 74.6 | 76.6 | 73.7 | 78.2 | 79.1 | 82.0 |

| Multi-Challenge | 63.8 | 42.5 | 63.0 | 63.1 | 63.9 | 64.1 | 63.5 |

| Long Context | | | | | | | |

| AA-LCR | 65.4 | 69.8 | 66.9 | 70.2 | 68.3 | 67.3 | 62.7 |

| RULER (1M) | 94.7 | -- | -- | -- | 90.1 | 94.2 | 87.7 |

| Longbench v2 (≤ 1M) | 61.9 | -- | -- | -- | 68.9 | 62.1 | 57.0 |

| Multilingual | | | | | | | |

| MMLU-ProX (avg en/de/fr/es/it/ja/zh/hi/pt/ko) | 83.0 | 78.4 | 85.8 | 85.0 | 86.4 | 85.6 | 84.3 |

| WMT24++ (en→xx) | 83.7 | 82.8 | 84.4 | 84.5 | 86.8 | 85.9 | 85.9 |

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, 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_v3 \
  --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_v3 \
  --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:v0.5.12.post1

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:v0.5.12.post1 \
  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_v3 \
  --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.

TensorRT-LLM#

Important Note: Current support for Nemotron 3 Ultra is limited to NVIDIA Blackwell architecture (including B200/B300 and GB200/GB300). While support for NVIDIA Hopper systems is planned, it is not currently available.

Container:

docker pull nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc17

For more detailed information, please see this cookbook.

8xB200:

cat > ./extra-llm-api-config.yml << EOF
backend: pytorch
trust_remote_code: true
tensor_parallel_size: 8
pipeline_parallel_size: 1
context_parallel_size: 1
gpus_per_node: 8
moe_expert_parallel_size: 1
disable_overlap_scheduler: false

cuda_graph_config:
  enable_padding: true
  max_batch_size: 256

enable_chunked_prefill: true
enable_attention_dp: false
max_batch_size: 256
max_seq_len: null
max_num_tokens: 32768
num_postprocess_workers: 4

kv_cache_config:
  enable_block_reuse: false
  max_tokens: null
  max_attention_window: null
  sink_token_length: null
  free_gpu_memory_fraction: 0.75
  host_cache_size: null
  cross_kv_cache_fraction: null
  secondary_offload_min_priority: null
  event_buffer_max_size: 0
  attention_dp_events_gather_period_ms: 5
  enable_partial_reuse: true
  copy_on_partial_reuse: true
  use_uvm: false
  max_gpu_total_bytes: 0
  iteration_stats_interval: 1
  dtype: fp8
  tokens_per_block: 32
  mamba_state_cache_interval: 256
  use_kv_cache_manager_v2: false
  max_util_for_resume: 0.95

moe_config:
  backend: TRTLLM
  max_num_tokens: null
  load_balancer: null
  disable_finalize_fusion: false
  use_low_precision_moe_combine: false
EOF


TLLM_ALLOW_LONG_MAX_MODEL_LEN=1 trtllm-serve  \
<bf16_ckpt> \
--max_batch_size 256 \
--tp_size 8 --ep_size 1 \
--max_num_tokens 32768 \
--trust_remote_code \
--reasoning_parser nano-v3 \
--tool_parser qwen3_coder \
--chat_template $MODEL_DIR/chat_template.jinja \
--extra_llm_api_options extra-llm-api-config.yml

Long-context configuration:

For long-context benchmarking, set TLLM_ALLOW_LONG_MAX_MODEL_LEN=1 as an environment variable and add --max_seq_len <seq_len> as the desired maximum context length. The MTP speculative_config block above carries over unchanged — on rc16, max_draft_len is the authoritative field and num_nextn_predict_layers is treated as deprecated.

API Client#

The examples below use the OpenAI-compatible client and work with any of the serving backends above.

NOTE: For coding agents add the following to the API call \- extra_body={"chat_template_kwargs": {"force_nonempty_content": True}}

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")
MODEL = "nvidia/nemotron-3-ultra"

Reasoning ON (default)

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "Write a haiku about GPUs"}],
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}}
)
print(response.choices[0].message.content)

Reasoning OFF

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What is the capital of Japan?"}],
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": False}}
)
print(response.choices[0].message.content)

Medium-effort reasoning

Uses significantly fewer reasoning tokens than full thinking mode. Recommended as a starting point before tuning explicit token budgets.

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What is the capital of Japan?"}],
    max_tokens=16000,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": True, "medium_effort": True}}
)
print(response.choices[0].message.content)

Tool calling with reasoning (SGLang requires explicit chat template kwargs)

response = client.chat.completions.create(
    model=MODEL,
    messages=[{"role": "user", "content": "What's the weather in New York?"}],
    tools=[{
        "type": "function",
        "function": {
            "name": "get_weather",
            "description": "Get the current weather for a city.",
            "parameters": {
                "type": "object",
                "properties": {
                    "city": {"type": "string"},
                    "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]}
                },
                "required": ["city"]
            }
        }
    }],
    tool_choice="required",
    max_tokens=256,
    temperature=1.0,
    top_p=0.95,
    extra_body={"chat_template_kwargs": {"enable_thinking": True, "force_nonempty_content": True}}
)

OpenCode#

OpenCode is an AI coding agent that runs in your terminal. It connects to any OpenAI-compatible endpoint, making it compatible with all three serving backends above (vLLM, SGLang, and TensorRT-LLM).

Create or update your ~/.config/opencode/opencode.json:

{
    "$schema": "https://opencode.ai/config.json",
    "model": "local/nvidia-nemotron-3-ultra",
    "provider": {
        "local": {
            "npm": "@ai-sdk/openai-compatible",
            "name": "local_backend",
            "options": {
                "baseURL": "http://localhost:8000/v1",
                "apiKey": "EMPTY"
            },
            "models": {
                "nvidia-nemotron-3-ultra": {
                    "name": "nvidia/nemotron-3-ultra",
                    "limit": {
                        "context": 1000000,
                        "output": 32768
                    }
                }
            }
        }
    },
    "agent": {
        "build": {
            "temperature": 1.0,
            "top_p": 0.95,
            "max_tokens": 32000
        },
        "plan": {
            "temperature": 1.0,
            "top_p": 0.95,
            "max_tokens": 32000
        }
    }
}

All backends above default to port 8000, so the baseURL works as-is for vLLM, SGLang, and TensorRT-LLM.

To learn more about other supported agent scaffolds \- check out this resource

Advanced: Budget-Controlled Reasoning

Set a hard token ceiling on the reasoning trace using reasoning_budget. The model will attempt to close the trace at the next newline before the budget is hit; if none is found within 500 tokens it closes abruptly at reasoning_budget + 500.

from typing import Any, Dict, List
import openai
from transformers import AutoTokenizer


class ThinkingBudgetClient:
    def __init__(self, base_url: str, api_key: str, tokenizer_name_or_path: str):
        self.tokenizer = AutoTokenizer.from_pretrained(tokenizer_name_or_path)
        self.client = openai.OpenAI(base_url=base_url, api_key=api_key)

    def chat_completion(
        self,
        model: str,
        messages: List[Dict[str, Any]],
        reasoning_budget: int = 512,
        max_tokens: int = 1024,
        **kwargs,
    ) -> Dict[str, Any]:
        assert max_tokens > reasoning_budget, (
            f"reasoning_budget must be less than max_tokens. "
            f"Got {max_tokens=} and {reasoning_budget=}"
        )

        # Step 1: generate the reasoning trace up to the budget
        response = self.client.chat.completions.create(
            model=model, messages=messages, max_tokens=reasoning_budget, **kwargs
        )
        reasoning_content = response.choices[0].message.content
        if "</think>" not in reasoning_content:
            reasoning_content = f"{reasoning_content}.\n\n</think>\n"

        reasoning_tokens_len = len(
            self.tokenizer.encode(reasoning_content, add_special_tokens=False)
        )
        remaining_tokens = max_tokens - reasoning_tokens_len
        assert remaining_tokens > 0, (
            f"No tokens remaining for response ({remaining_tokens=}). "
            "Increase max_tokens or lower reasoning_budget."
        )

        # Step 2: continue from the closed reasoning trace
        messages.append({"role": "assistant", "content": reasoning_content})
        prompt = self.tokenizer.apply_chat_template(
            messages, tokenize=False, continue_final_message=True
        )
        response = self.client.completions.create(
            model=model, prompt=prompt, max_tokens=remaining_tokens, **kwargs
        )

        return {
            "reasoning_content": reasoning_content.strip().strip("</think>").strip(),
            "content": response.choices[0].text,
            "finish_reason": response.choices[0].finish_reason,
        }

Example usage (32-token reasoning budget):

client = ThinkingBudgetClient(
    base_url="http://localhost:8000/v1",
    api_key="EMPTY",
    tokenizer_name_or_path="nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16",
)

result = client.chat_completion(
    model="nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16",
    messages=[
        {"role": "system", "content": "You are a helpful assistant. /think"},
        {"role": "user", "content": "What is 2+2?"},
    ],
    reasoning_budget=32,
    max_tokens=512,
    temperature=1.0,
    top_p=0.95,
)
print(result)

Training and Evaluation Datasets#

Training#

Data Modality: Text The total size: 53.8 TiB (14.8 trillion tokens) Total number of datasets: 226 Dataset partition: Training [100%], testing [0%], validation [0%] Time period for training data collection: 2013 to 2026 Time period for testing data collection: 2013 to 2026 Time period for validation data collection: 2013 to 2026 Data Collection Method by dataset: Hybrid: Automated, Human, Synthetic Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 is pre-trained on a large corpus of high-quality curated and synthetically-generated data. It is trained in the English language, as well as 11 other languages and 43 programming languages. Our sources cover a variety of document types such as: webpages, dialogue, articles, and other written materials. The corpus spans domains including legal, math, science, finance, and more. We also include a small portion of question-answering, and alignment style data to improve model accuracy. The model was pre-trained for approximately 20 trillion tokens.

The post-training corpus for NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16 consists of high-quality curated and synthetically-generated data. Primary languages used for post-training include English, French, Spanish, Italian, German, Japanese, Hindi, Korean, Brazilian Portuguese, and Chinese.

These datasets, such as FinePDFs, EssentialWeb, HotpotQA, SQuAD, and HelpSteer3, do not collectively or exhaustively represent all demographic groups (and proportionally therein). For instance, these datasets do not contain explicit mentions of demographic classes such as age, gender, or ethnicity in 64-99% of samples, depending on the source. In the subset where such terms are present, document-based datasets (FinePDFs and EssentialWeb) contain representational skews, such as references to "male" outnumbering those to "female", and mentions of "White" as the most frequent among ethnic identifiers (comprising 43-44% of ethnicity mentions). To mitigate these imbalances, we recommend considering evaluation techniques such as bias audits, fine-tuning with demographically balanced datasets, and mitigation strategies like counterfactual data augmentation to align with the desired model behavior. This evaluation used a 3,000-sample subset per dataset, identified as the optimal threshold for maximizing embedder accuracy.

During post-training, we generate synthetic data by distilling trajectories, solutions, and translations from strong teacher models and agent systems, often grounded in real tasks or documents and aggressively filtered for quality. For math, code, and science, we start from curated problem sets and use open source permissive models such as GPT-OSS-120B to produce step-by-step reasoning traces, candidate solutions, best-of-n selection traces, and verified CUDA kernels. For long-context and science, we build synthetic QA and reasoning data by retrieving passages from long documents, generating MCQ/OpenQA questions and answers, and paraphrasing them into multiple prompt/response formats to ensure diversity. Across all pipelines we stack automated verification—compilers, numerical checks, language identification—to ensure our data is high quality.

For all domains, we apply a unified data filtering pipeline to ensure that only high-quality, license-compliant, and verifiable samples are used for post-training. We first discard malformed examples using structural checks (e.g., missing tool definitions when tool calls are present). We then aggressively filter reasoning traces exhibiting pathological repetition, such as repeated n-grams within a sliding window or across the entire trajectory, which we found to be a strong indicator of malformed or low-quality reasoning. Finally, based on internal audits of synthetically generated datasets, we observed that some teacher models occasionally produce reasoning traces and final responses that implicitly align with specific political entities or promote nationalistic narratives. To mitigate this, we apply targeted keyword- and regex-based filters and remove all trajectories matching such behavior.

Alongside the model, we release our final pre-training and post-training data, as outlined in this section. For ease of analysis, there is a sample set that is ungated. For all remaining code, math and multilingual data, gating and approval is required, and the dataset is permissively licensed for model training purposes.

More details on the datasets and synthetic data generation methods can be found in the technical report NVIDIA Nemotron 3 Ultra.

For more information about the datasets used to train this model, please see the Public Summary of Training Content

For Detailed Dataset Information: Click here!

Base Pre-Training Corpus (Nemotron 3 Foundation)#

The foundation of the model is trained on the Nemotron-3-Ultra corpus, comprising the following datasets from the Nemotron Pretraining Datasets collection:

Public Datasets#

Crawled and Scraped from Online Sources by NVIDIA#

The English Common Crawl data was downloaded from the Common Crawl Foundation (see their FAQ for details on their crawling) and includes the snapshots CC-MAIN-2013-20 through CC-MAIN-2025-13. The data was subsequently deduplicated and filtered in various ways described in the Nemotron-CC paper. Additionally, we extracted data for fifteen languages from the following three Common Crawl snapshots: CC-MAIN-2024-51, CC-MAIN-2025-08, CC-MAIN-2025-18. The fifteen languages included were Arabic, Chinese, Danish, Dutch, French, German, Italian, Japanese, Korean, Polish, Portuguese, Russian, Spanish, Swedish, and Thai. As we did not have reliable multilingual model-based quality classifiers available, we applied just heuristic filtering instead—similar to what we did for lower quality English data in the Nemotron-CC pipeline, but selectively removing some filters for some languages that did not work well. Deduplication was done in the same way as for Nemotron-CC.

The GitHub Crawl was collected using the GitHub REST API and the Amazon S3 API. Each crawl was operated in accordance with the rate limits set by its respective source, either GitHub or S3. We collect raw source code and subsequently remove any having a license which does not exist in our permissive-license set (for additional details, refer to the technical report).

Private Non-publicly Accessible Datasets of Third Parties#

Private Non-publicly Accessible Datasets by NVIDIA#

NVIDIA-Sourced Synthetic Datasets (Pre-Training)#

The publisher's README is longer; it is cut here. Read the rest at the source.