Model Overview

Description:

NVIDIA Nemotron 3 Nano Omni is a multimodal large language model that unifies video, audio, image, and text understanding to support enterprise-grade Q&A, summarization, transcription, and document intelligence workflows. It extends the Nemotron Nano family with integrated video+speech comprehension, Graphical User Interface (GUI), Optical Character Recognition (OCR), and speech transcription capabilities, enabling end-to-end processing of rich enterprise content such as meeting recordings, M&E assets, training videos, and complex business documents. NVIDIA Nemotron 3 Nano Omni was developed by NVIDIA as part of the Nemotron model family.

This model is available for commercial use.

This model was improved using Qwen3-VL-30B-A3B-Instruct, Qwen3.5-122B-A10B, Qwen3.5-397B-A17B, Qwen2.5-VL-72B-Instruct, and gpt-oss-120b. For more information, please see the Training Dataset section below.

License/Terms of Use

Governing Terms: This trial service is governed by the NVIDIA API Trial Terms of Service. Use of this model is governed by the NVIDIA Open Model Agreement.

Deployment Geography:

Global

Use Case:

This model is designed for enterprise customers requiring multimodal understanding capabilities. Expected users include:

• Customer service applications (e.g., Doordash video of drop-off at a given address via OCR, drive-thru order verification)
• Media and Entertainment (M&E) — video and speech analysis, dense captions, video search and summarization
• Document intelligence for AI assistants (contracts, SOW/MSA, scientific discovery, financial documents)
• GUI automation for AI agentic applications (incident management, agentic search, browser agents, email agents)

Release Date:

Build.Nvidia.com 04/28/2026 via URL
Hugging Face 04/28/2026 via:

• BF16
  
• FP8
  
• NVFP4
  

NGC 04/28/2026 via URL

Model Architecture:

Architecture Type: Mamba2-Transformer Hybrid Mixture of Experts (MoE)

Network Architecture:

• Nemotron 3 Nano LLM (30B A3B)
• CRADIO v4-H vision encoder
• Parakeet speech encoder

Number of model parameters: 3.1 x 10^10 (31B A3B)

Input(s):

Input Type(s): Video, Audio, Image, Text

Input Format(s):

• Video: mp4, up to 2 minutes. For 1080p videos, sample up to 1 FPS / 128 frames. For lower-resolution videos such as 720p, higher temporal sampling such as 2 FPS / 256 frames may be used.
• Audio: wav, mp3 files (up to 1 hour), 8kHz and higher sampling rates
• Image: Red, Green, Blue (RGB) (jpeg, png)
• Text: String
  

Input Parameters:

• Video: Three-Dimensional (3D)
• Audio: One-Dimensional (1D)
• Image: Two-Dimensional (2D)
• Text: One-Dimensional (1D)
  

Other Properties Related to Input:

• Maximum context length up to 256k tokens
• Language support: English only
  

Output(s)

Output Type(s): Text

Output Format(s):

• Text: String
  

Output Parameters:

• Text: One-Dimensional (1D)
  

Other Properties Related to Output:

• Maximum context length up to 256k tokens.
• Supports JSON output format
• Supports reasoning output with chain-of-thought
• Supports tool calling
• Supports word-level timestamps for transcription
  

Our AI models are designed and/or 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):

• vLLM
  
• NeMo
  
• Megatron
  
• NeMo-RL
  

Supported Hardware Microarchitecture Compatibility:

• NVIDIA Ampere (A100 80GB SXM/NVLink)
  
• NVIDIA Blackwell (B200 SXM/NVLink, RTX Pro 6000 SE, DGX Spark, Jetson Thor, RTX 5090)
  
• NVIDIA Hopper (H100 SXM/NVLink, H200 SXM/NVLink)
  
• NVIDIA Lovelace (L40S)
  

Preferred/Supported Operating System(s):

• Linux
  

Inference Runtimes:

• vLLM
  
• TensorRT LLM
  
• TensorRT Edge-LLM
  
• llama.cpp
  
• Ollama
  
• SGLang
  

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.

This AI model can be embedded as an Application Programming Interface (API) call into the software environment described above.

Model Version(s):

Nemotron-3-Nano-Omni-30B-A3B-Reasoning

Quick Start Guide

Model Parameters

Download Model Weights

Install the HuggingFace CLI

pip install -U "huggingface_hub[hf_xet]"

# Log in once; the token is cached at ~/.cache/huggingface/token
hf auth login

# Sanity check: should print your username and orgs
hf auth whoami

Download the weights

Pick a target directory on a volume with ≥70 GB free (the model is ~62 GB).

WEIGHTS=/path/to/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16

hf download nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 \
  --local-dir "$WEIGHTS" \
  --max-workers 8

Notes:

• hf download is resumable — re-run the same command if the connection drops.
• --max-workers 8 parallelizes downloads; tune up on fast networks.
• The hf_xet extra enables native Xet-protocol transfers for Xet-backed repos; no need for git-xet or git-lfs when using hf download.

Verify the download

ls "$WEIGHTS" | head
du -sh "$WEIGHTS"  # expect ~62 GB
test -f "$WEIGHTS/config.json" && echo OK

vLLM

Required version: vLLM 0.20.0 is needed. This means one of these containers:

• CUDA 13.0: 'vllm/vllm-openai.20.0'
• CUDA 12.9: 'vllm/vllm-openai.20.0-cu129'

Container

docker pull vllm/vllm-openai:v0.20.0

Audio support: Within the vLLM container, before running vllm serve, if any audio will be used (including passing use_audio_in_video: true):

Bash

CopiedCopy

python3 -m pip install "vllm[audio]"

General Invocation (1×GPU, e.g. 1×B200)

# vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 \
# vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-FP8 \
vllm serve nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4 \
  --host 0.0.0.0 \
  --max-model-len 131072 \
  --tensor-parallel-size 1 \
  --trust-remote-code \
  --video-pruning-rate 0.5 \
  --max-num-seqs 384 \
  --allowed-local-media-path / \
  --media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}' \
  --reasoning-parser nemotron_v3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder \
  --kv-cache-dtype fp8 # Omit this for BF16

Platform-Specific Notes

RTX Pro: Due to a current bug with FlashInfer + RTX Pro, append: --moe-backend triton

NVFP4 + TP>1: Due to a current bug with the TRTLLM_GEN MoE backend kernels on vLLM, when running with TP>1 on NVFP4, append: --moe-backend flashinfer_cutlass

vLLM on DGX Spark (aarch64 / ARM64)

For everything not covered here (API examples, reasoning mode, video tuning), follow the general instructions.

1. Pull the container image

Use the upstream multi-arch vLLM v0.20.0 docker image. Docker will automatically pull the arm64 variant.

docker pull vllm/vllm-openai:v0.20.0

2. Launch the vLLM server on Spark

WEIGHTS=/path/to/nemotron-3-nano-omni-weights

# The image does not include audio packages so we need to install them with "pip install vllm[audio]" as done in the command below
docker run --rm -it \
  --gpus all \
  --ipc=host -p 8000:8000 \
  --shm-size=16g \
  --name vllm-nemotron-omni \
  -v "${WEIGHTS}:/model:ro" \
  --entrypoint /bin/bash \
  vllm/vllm-openai:v0.20.0 -c  \
  "pip install vllm[audio] && vllm serve /model \
  --served-model-name=nemotron_3_nano_omni \
  --max-num-seqs 8 \
  --max-model-len 131072 \
  --port 8000 \
  --trust-remote-code \
  --gpu-memory-utilization 0.8 \
  --limit-mm-per-prompt '{\"video\": 1, \"image\": 1, \"audio\": 1}' \
  --media-io-kwargs '{\"video\": {\"fps\": 2,  \"num_frames\": 256}}' \
  --allowed-local-media-path=/ \
  --enable-prefix-caching \
  --max-num-batched-tokens 32768 \
  --reasoning-parser nemotron_v3 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder"

In another terminal, verify the server is ready:

curl -sS http://localhost:8000/v1/models | python3 -m json.tool

Key Spark-Specific Flags

Memory Tuning on Spark

Spark uses unified LPDDR5X memory (~128 GB shared between CPU and GPU), not separate system + VRAM pools. Two levers, in order of impact:

1. Lower --gpu-memory-utilization from 0.85 → 0.70 to free ~19 GB back to the OS and re-enable weight prefetch. Cost: smaller KV cache budget.
2. Lower --max-model-len to reduce KV cache allocation (e.g. halving context window halves KV cache at --max-num-seqs=1). Combined override:

  --gpu-memory-utilization=0.70 \
  --max-model-len=32768 \

TensorRT-LLM

This model can also be deployed with TensorRT-LLM - see relevant instructions here.

Platform-Specific Notes

TensorRT Edge-LLM

This model can also be deployed with TensorRT Edge-LLM on NVIDIA Jetson Thor - see the Jetson AI Lab model page and the TensorRT Edge-LLM Quick Start Guide.

SGLang

The BF16 variant of this model is supported on SGLang, with the following images:

• CUDA 13.0: lmsysorg/sglang:dev-cu13-nemotronh-nano-omni-reasoning-v3
• CUDA 12.9: lmsysorg/sglang:dev-nemotronh-nano-omni-reasoning-v3

librosa must be installed first: pip install librosa --break-system-packages

To serve: sglang serve --model-path nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 --trust-remote-code

Platform-Specific Notes

SGLang on DGX Spark (aarch64 / ARM64)

For everything not covered here (API examples, reasoning mode, video tuning), follow the general instructions.

1. Pull the container image

Use the upstream multi-arch CUDA 13.0 docker image linked above. Docker will automatically pull the arm64 variant.

docker pull lmsysorg/sglang:dev-cu13-nemotronh-nano-omni-reasoning-v3

2. Launch the SGLang server on Spark

WEIGHTS=/path/to/nemotron-3-nano-omni-weights

# The image does not include audio packages so we need to install them with "pip install librosa" as done in the command below
docker run --gpus all -it --rm \
  -p 30000:30000 \
  -v "${WEIGHTS}:/model:ro" \
  --shm-size 16g \
  lmsysorg/sglang:dev-cu13-nemotronh-nano-omni-reasoning-v3 \
  bash -c "pip install librosa && python3 -m sglang.launch_server --model-path /model \
  --host 0.0.0.0 \
  --port 30000 \
  --trust-remote-code \
  --mem-fraction-static 0.8 \
  --max-running-requests 8 \
  --tool-call-parser qwen3_coder \
  --reasoning-parser nemotron_3"

In another terminal, verify the server is ready:

curl -sS http://localhost:30000/v1/models | python3 -m json.tool

Key Spark-Specific Flags

Memory Tuning on Spark

Spark uses unified LPDDR5X memory (~128 GB shared between CPU and GPU), not separate system + VRAM pools. Two levers, in order of impact:

1. Lower --mem-fraction-static from 0.80 → 0.70 to free ~13 GB back to the OS and re-enable weight prefetch. Cost: smaller KV cache budget.
2. Lower --context-length to reduce KV cache allocation (e.g. halving context window halves KV cache at --max-running-requests=1). Combined override:

  --mem-fraction-static=0.70 \
  --context-length=32768 \

API Client (OpenAI-compatible)

from openai import OpenAI
client = OpenAI(base_url="http://localhost:8000/v1", api_key="")
MODEL = "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4"

Image Example

import base64

def image_to_data_url(path: str) -> str:
    with open(path, "rb") as f:
        b64 = base64.b64encode(f.read()).decode("utf-8")
    return f"data:image/jpeg;base64,{b64}"

image_url = image_to_data_url("media/example1a.jpeg")

response = client.chat.completions.create(
    model=MODEL,
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Describe this image in detail."},
                {"type": "image_url", "image_url": {"url": image_url}},
            ],
        }
    ],
    max_tokens=1024,
    temperature=1.0,
    extra_body={"top_k": 1, "chat_template_kwargs": {"enable_thinking": False}},
)
print(response.choices[0].message.content)

Audio Example

from pathlib import Path

audio_url = Path("media/2414-165385-0000.wav").resolve().as_uri()

response = client.chat.completions.create(
    model=MODEL,
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "audio_url", "audio_url": {"url": audio_url}},
                {"type": "text", "text": "Transcribe this audio."},
            ],
        }
    ],
    max_tokens=1024,
    temperature=1.0,
    extra_body={"top_k": 1, "chat_template_kwargs": {"enable_thinking": False}},
)
print(response.choices[0].message.content)

Video Example

from pathlib import Path

video_url = Path("media/demo.mp4").resolve().as_uri()
reasoning_budget = 16384
grace_period = 1024

response = client.chat.completions.create(
    model=MODEL,
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "video_url", "video_url": {"url": video_url}},
                {"type": "text", "text": "Describe this video."},
            ],
        }
    ],
    max_tokens=20480,
    temperature=0.6,
    top_p=0.95,
    extra_body={
        "thinking_token_budget": reasoning_budget + grace_period,
        "chat_template_kwargs": {
            "enable_thinking": True,
            "reasoning_budget": reasoning_budget,
        },
        "mm_processor_kwargs": {"use_audio_in_video": False},
    },
)
print(response.choices[0].message.content)

Text Example (curl)

curl -sS http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4","messages":[{"role":"user","content":"Hello, what can you do?"}],"temperature":1.0,"top_k":1}' \
  | python3 -c "import sys,json; print(json.load(sys.stdin)['choices'][0]['message']['content'])"

PDF Example (page-by-page via Python)

The API accepts images, not raw PDF files. The script below renders each page to PNG and sends it as base64. Save as pdf_vlm_chat.py and install dependencies: pip install pymupdf pillow requests.

#!/usr/bin/env python3
"""Send PDF page(s) as images to a vLLM /v1/chat/completions endpoint."""
from __future__ import annotations

import argparse, base64, sys
from io import BytesIO
from pathlib import Path

import requests

try:
    import fitz
    from PIL import Image
except ImportError:
    print("Install: pip install pymupdf pillow requests", file=sys.stderr)
    sys.exit(1)

USER_PROMPT = (
    "Summarize this PDF page: main topic, section headings, important facts "
    "or bullets, and a brief note on each figure or table. "
    "Do not invent text you cannot read."
)
API_URL = "http://localhost:8000/v1/chat/completions"
MODEL = "nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-NVFP4"
MAX_TOKENS = 32000
DPI = 150


def page_to_b64(pdf_path: str, idx: int) -> str:
    doc = fitz.open(pdf_path)
    z = DPI / 72.0
    pix = doc.load_page(idx).get_pixmap(matrix=fitz.Matrix(z, z))
    img = Image.frombytes("RGB", [pix.width, pix.height], pix.samples)
    doc.close()
    buf = BytesIO()
    img.save(buf, format="PNG")
    return base64.b64encode(buf.getvalue()).decode("ascii")


def chat(url, model, b64, text, max_tokens):
    r = requests.post(url, json={
        "model": model,
        "messages": [{"role": "user", "content": [
            {"type": "text", "text": text},
            {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{b64}"}},
        ]}],
        "max_tokens": max_tokens,
        "stream": False,
        "temperature": 1.0,
        "chat_template_kwargs": {"enable_thinking": False},
    }, timeout=120)
    r.raise_for_status()
    return r.json()["choices"][0]["message"]["content"]


def main():
    p = argparse.ArgumentParser()
    p.add_argument("pdf")
    p.add_argument("--page", type=int, default=0)
    p.add_argument("--all-pages", action="store_true")
    p.add_argument("-o", "--output")
    p.add_argument("--url", default=API_URL)
    p.add_argument("--model", default=MODEL)
    p.add_argument("--max-tokens", type=int, default=MAX_TOKENS)
    a = p.parse_args()

    doc = fitz.open(a.pdf); n = len(doc); doc.close()
    pages = range(n) if a.all_pages else [a.page]
    parts = [f"# Extracted: {Path(a.pdf).name}\n\n*Pages: {n}*\n"] if a.all_pages else []

    for i in pages:
        print(f"Page {i+1}/{n} ...", file=sys.stderr)
        b64 = page_to_b64(a.pdf, i)
        text = chat(a.url, a.model, b64, f"Page {i+1}.\n\n{USER_PROMPT}", a.max_tokens)
        parts.append(f"\n---\n\n## Page {i+1}\n\n{text.strip()}\n" if a.all_pages else text.strip())

    out = "\n".join(parts)
    if a.output:
        Path(a.output).write_text(out + "\n", encoding="utf-8")
    else:
        print(out)

if __name__ == "__main__":
    main()

Single page:

python3 pdf_vlm_chat.py /path/to/your_document.pdf --page 0

All pages to markdown:

python3 pdf_vlm_chat.py /path/to/your_document.pdf --all-pages -o extracted.md

Edit USER_PROMPT in the script for different tasks (detailed extraction, table parsing, etc.).

Reasoning Mode (enable_thinking)

To disable reasoning on a request, add to the JSON body:

"chat_template_kwargs": {"enable_thinking": false}

In the Python heredoc pattern, use False (Python boolean), not false (invalid Python).

We recommend thinking mode for tasks that involve reasoning and complex understanding. For video, audio, and omni use cases, try both enabling and disabling thinking for best results.

from typing import Any, Dict, List

from openai 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, trust_remote_code=True
        )
        self.client = 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 only the reasoning trace up to the requested budget.
        response = self.client.chat.completions.create(
            model=model,
            messages=messages,
            max_tokens=reasoning_budget,
            extra_body={
                "top_k": 1,
                "chat_template_kwargs": {
                    "enable_thinking": True,
                },
            },
            **kwargs,
        )
        reasoning_content = response.choices[0].message.content or ""
        if "</think>" not in reasoning_content:
            print("No </think> found in reasoning content")
            reasoning_content = f"{reasoning_content}</think>\n\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 and ask for the final answer.
        continued_messages = messages + [
            {"role": "assistant", "content": reasoning_content}
        ]
        prompt = self.tokenizer.apply_chat_template(
            continued_messages,
            tokenize=False,
            continue_final_message=True,
        )
        response = self.client.completions.create(
            model=model,
            prompt=prompt,
            max_tokens=remaining_tokens,
            extra_body={"top_k": 1},
            **kwargs,
        )

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

Video Tuning

Frame sampling (--media-io-kwargs)

Without explicit settings, vLLM may default to ~32 frames per video regardless of length. Always set --media-io-kwargs at server launch (already included in the General Invocation above):

--media-io-kwargs '{"video": {"fps": 2, "num_frames": 256}}'

Recommended num_frames ranges (at fps=2):

Higher values improve temporal coverage but increase VRAM and prefill time. Start at the low end of the range and increase as your workload and latency budget allow.

Notes

1. Reasoning default: Reasoning is on by default. If you omit chat_template_kwargs, the model will produce chain-of-thought traces in content. This is appropriate for text and image inputs.
2. Video frame sampling: The default (~32 frames) is too conservative for most real videos. Set --media-io-kwargs at server launch.
3. PDF input format: The API does not accept raw PDF uploads. Render pages to PNG and send as base64 (see PDF Example above).
4. max_tokens vs --max-model-len: max_tokens in the request caps only the completion (generated output). It cannot exceed the server's --max-model-len, which is the hard ceiling for prompt + completion combined. Increase the server flag if you need longer outputs.

Jetson Deployment

For Jetson deployments, vLLM, SGLang, Ollama, llama.cpp, and TensorRT Edge-LLM are supported inference frameworks; see the Jetson AI Lab model page for more details.

TensorRT Edge-LLM support is only for Jetson Thor; TensorRT-LLM is not supported on Jetson.

Training, Testing, and Evaluation Datasets:

Dataset Overview

Total Size: 354,587,705 data points (~717.0B tokens)
Total Number of Datasets: 1395 dataset entries

Dataset partition: Training [100%], Testing [N/A — evaluation benchmarks used separately], Validation [N/A — evaluation benchmarks used separately]
Time period for training data collection: 2019–2025
Time period for testing data collection: N/A (standard public benchmarks)
Time period for validation data collection: N/A (standard public benchmarks)

Dataset Description

Nemotron-Omni extends our commitment from text to multimodal, delivering the same level of openness across text, audio, image, and video.

Adapter and encoder training scale: ~127B tokens across mixed modalities spanning text+image, text+video, text+audio, and text+video+audio—reflecting real-world, contextualized interactions versus single-modality data.

Post-training for real-world tasks: ~124M curated examples across multimodal combinations (text+audio, text+image, text+video, and text+video+audio), structured to support document reasoning, computer use, and long-horizon workflows.

RL environments for agent training: 20 RL datasets across 25 environments covering 5 new multimodal tasks—visual grounding, chart and document understanding, vision-critical STEM problems, video understanding, and automatic speech recognition—extending Nemotron's RL pipeline beyond text into vision and audio.

Modality Breakdown:

Training data for Nemotron-Omni was assembled from a diverse collection of audio, image, video, and text datasets. Raw datasets were first converted into a standardized JSONL format with unified conversation-turn structure. Audio data was resampled to 16 kHz where needed. Image and video datasets were paired with question-answer annotations, often regenerated or refined using large vision-language models to improve quality and consistency. Quality filtering was applied using model-based judges to remove low-quality, unsafe, or off-topic samples. Deduplication and CSAM scanning were performed across all image datasets. Data was then packed into fixed-length sequences (32k, 128k, or 256k tokens) for efficient training.

Multiple safety measures were implemented throughout the data pipeline. All image/text datasets underwent CSAM (Child Sexual Abuse Material) scanning, with results tracked per dataset. Content safety filtering was applied using two independent safety judge models to flag and remove samples containing harmful content including weapons references, criminal planning, sexual content involving minors, harassment, hate speech, profanity, threats, violence, or suicide-related content. Synthetic data generation pipelines included explicit quality and safety filtering stages. Identity-fix processing was applied to correct potential biases in generated responses. The multi-stage pipeline (original → cleaned → clean+safe → clean+safe+holdout) ensured progressive refinement, with each stage removing additional problematic content.

We built on the base model, applying additional training, enhancements, and optimizations on top of it.

Public Datasets

Private Datasets

Self-Sourced Synthetic Data

• Overall Size: 41,502,625 samples across modalities: text+audio, text+image, text+video
  

• Description of synthetic data generation methods:
  

Synthetic data generation (SDG) was used to improve data quality, generate reasoning traces, re-label annotations, and augment existing datasets. Methods include: re-captioning images and audio using vision-language models, generating question-answer pairs from existing media, producing thinking/reasoning chains for complex tasks, paraphrasing prompts for diversity, and applying model-based quality filtering.

NVIDIA-Sourced Synthetic Datasets

Training Dataset:

Data Modality

• Audio
  
• Image
  
• Text
  
• Video
  

Audio Training Data Size

• 10,000 to 1 Million Hours
  (267,898,865 audio-containing samples)
  

Image Training Data Size

• 1 Million to 1 Billion Images
  (70,143,901 image-containing samples)
  

Text Training Data Size

• 1 Billion to 10 Trillion Tokens
  (~717.0B tokens total across all modalities)
  

Video Training Data Size

• 10,000 to 1 Million Hours
  (24,557,717 video-containing samples)
  

Data Collection Method by dataset

• Hybrid: Human, Automated, Synthetic
  

Labeling Method by dataset

• Hybrid: Human, Automated, Synthetic
  

Properties (Quantity, Dataset Descriptions, Sensor(s)): 354,587,705 total data items across 1395 datasets. The training data spans five modality combinations: text+audio (259,178,821 samples), text+image (70,143,901 samples), text+video (15,837,673 samples), text+video+audio (8,720,044 samples), and text-only (707,187 samples). Content includes publicly available academic datasets, licensed third-party data, NVIDIA-internal collections, and synthetically generated annotations. The data is primarily in English. No sensor-derived data was used.

Evaluation Dataset:

Benchmark Scores:

Quantization Benchmark Scores:

We release FP8 and NVFP4 quantized variants alongside the BF16 model. The FP8 variant quantizes every linear layer in the language model to per-tensor E4M3 (with the exception of the MoE router and lm_head) and pairs it with an FP8 KV cache, yielding 8.5 effective bits per weight (32.8 GB). The NVFP4 variant uses a mixed-precision recipe inspired by Nemotron 3 Super: routed MoE experts are quantized to NVFP4 (FP4 E2M1 values with per-block FP8 E4M3 scales over groups of 16 elements and an additional per-tensor FP32 global scale), while the Mamba in_proj / out_proj, shared experts, and attention o_proj are quantized to FP8, yielding 4.98 effective bits per weight (20.9 GB). In both variants the vision and audio encoders and their MLP projectors are kept in BF16.

The table below reports FP8 & NVFP4 accuracy against a BF16 baseline using non-reasoning mode. Across 9 multimodal benchmarks, both quantized variants stay within 1 point of BF16 on average.

Data Collection Method by dataset:

• Hybrid: Human, Automated — Evaluation benchmarks are primarily human-curated public academic datasets with automated scoring.
  

Labeling Method by dataset:

• Human
  

Properties (Quantity, Dataset Descriptions, Sensor(s)): 14 evaluation benchmarks spanning image understanding (MathVistaMini, Charxiv Reasoning, MMLongBench-Doc, OCR Reasoning, OCRBenchV2 English, CVBench2D, OSWorld), video understanding (Video MME), audio/speech understanding (VoiceBench, Tedium Long, HF-ASR, MMAU, World Sense), and multimodal omni-understanding (Daily Omni). All benchmarks are publicly available academic datasets in English.

Prior to training this model, NVIDIA implemented measures to respect EU text and data mining opt-outs by (1) respecting robots.txt instructions to the extent such signals reflect valid rights reservations, and (2) filtering datasets on any actionable metadata identifiers provided by rightsholders.

Inference:

Acceleration Engine: vLLM
Test Software: vLLM
Test Hardware:

• NVIDIA H100 x2
  

Best Practices

We recommend following settings for reaching the optimal performance.

Sampling Parameters

We suggest the following sampling parameters based on the mode and tasks.

• Thinking mode for long document analysis and multimodal reasoning tasks:
  temperature=0.5-0.7, top_p=0.95, grace_period=1024, reasoning_budget=16384, max_token=20480, and max_model_len=210000
  
• Instruct mode (non-thinking) for general tasks:
  temperature=0.2, top_k=1
  
• For ASR tasks, we recommend non-thinking mode with temperature=0.2, top_k=1
  

Model output length

For most multimodel reasoning tasks, we recommend using output length of at least 20480. For complex reasoning questions especially in math and programing increasing the maximum output length to 131072 tokens can give the model enough room to produce more detailed and correct answers. We also found the proposed Budget-Controlled Reasoning effectiveness in answering complex reasoning questions.

Ethical Considerations:

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please make sure you have proper rights and permissions for all input image and video content; if image or video includes people, personal health information, or intellectual property, the image or video generated will not blur or maintain proportions of image subjects included.

For more detailed information on ethical considerations for this model, please see the Model Card++ Bias, Explainability, Safety & Security, and Privacy Subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.