Model Overview

Description:

NVIDIA Nemotron Nano v2 12B VL model enables multi-image reasoning and video understanding, along with strong document intelligence, visual Q&A and summarization capabilities.

This model is ready for commercial use.

License/Terms of Use

Governing Terms: Use of this model is governed by the NVIDIA Open Model License Agreement

Deployment Geography:

Global

Use Case:

Nemotron Nano 12B V2 VL is a model for multi-modal document intelligence. It would be used by individuals or businesses that need to process documents such as invoices, receipts, and manuals. The model is capable of handling multiple images of documents, up to four images at a resolution of 1k x 2k each, along with a long text prompt. The expected use is for tasks like summarization and Visual Question Answering (VQA). The model is also expected to have a significant advantage in throughput.

Release Date:

• Build.Nvidia.com [October 28th, 2025] via nvidia/NVIDIA-Nemotron-Nano-VL-12B-V2
• Hugging Face [October 28th, 2025] via nvidia/NVIDIA-Nemotron-Nano-VL-12B-V2-BF16
• Hugging Face [October 28th, 2025] via nvidia/NVIDIA-Nemotron-Nano-VL-12B-V2-FP8
• Hugging Face [October 28th, 2025] via nvidia/NVIDIA-Nemotron-Nano-VL-12B-V2-NVFP4

Model Architecture:

Architecture Type: Transformer
Network Architecture: Vision Encoder: CRadioV2-H Language Encoder: NVIDIA-Nemotron-Nano-12B-v2

** Number of model parameters: 12.6B

Computational Load (For NVIDIA Models Only)

Cumulative Compute: 2.2e+22
Estimated Energy and Emissions for Model Training: Energy Consumption: 7,827.46 kWh
Carbon Emissions: 3.21 tCO2e

Input:

Input Type(s): Image, Video, Text Input Format: Image (png,jpg), Video (MP4, MKV, FLV, 3GP), Text (String)
Input Parameters: Image (2D),Video(3D), Text (1D)
Other Properties Related to Input:

• Input Images Supported: 4
• Language Supported: English only
  
• Input + Output Token: 128K
• Minimum Resolution: 32 × 32 pixels
• Maximum Resolution: Determined by a 12-tile layout constraint, with each tile being 512 × 512 pixels. This supports aspect ratios such as:
  • 4 × 3 layout: up to 2048 × 1536 pixels
  • 3 × 4 layout: up to 1536 × 2048 pixels
  • 2 × 6 layout: up to 1024 × 3072 pixels
  • 6 × 2 layout: up to 3072 × 1024 pixels
  • Other configurations allowed, provided total tiles ≤ 12
• Channel Count: 3 channels (RGB)
• Alpha Channel: Not supported (no transparency)
  
• Frames: 2 FPS with min of 8 frame and max of 128 frames

Output:

Output Type(s): Text
Output Format: String
Output Parameters: 1D
Other Properties Related to Output: Input + Output Token: 128K

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]
  
• [TRT-LLM]
  
• [SGLang]
  

Supported Hardware Microarchitecture Compatibility:

• NVIDIA L40S
  *
• NVIDIA A100
  
• NVIDIA B200
  
• NVIDIA H100/H200
  
• NVIDIA RTX PRO 6000 Server Edition
  
• NVIDIA GB200
  

Preferred/Supported 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

Quick Start

Install Dependencies

pip install causal_conv1d "transformers>4.53,<4.54" torch timm "mamba-ssm==2.2.5" accelerate open_clip_torch numpy pillow

Usage

Single Image Inference

import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer

# Load model and processor
model_path = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"  # Or use a local path
device = "cuda:0"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map=device,
    torch_dtype=torch.bfloat16
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

# Load image
image = Image.open("path/to/your/image.jpg")

# Prepare messages
messages = [
    {"role": "system", "content": "/no_think"},
    {
        "role": "user",
        "content": [
            {"type": "image", "image": ""},
            {"type": "text", "text": "Describe the image."},
        ],
    }
]

# Generate prompt and process inputs
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[prompt], images=[image], return_tensors="pt").to(device)

# Generate output
generated_ids = model.generate(
    pixel_values=inputs.pixel_values,
    input_ids=inputs.input_ids,
    attention_mask=inputs.attention_mask,
    max_new_tokens=1024,
    do_sample=False,
    eos_token_id=tokenizer.eos_token_id,
)

# Decode output
output_text = processor.batch_decode(
    generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)[0]
print(output_text)

Multi-Image Inference

import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer

# Load model and processor
model_path = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"  # Or use a local path
device = "cuda:0"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map=device,
    torch_dtype=torch.bfloat16
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

# Load multiple images
images = [
    Image.open("path/to/image1.jpg"),
    Image.open("path/to/image2.jpg"),
]

# Prepare messages with multiple images
messages = [
    {"role": "system", "content": "/no_think"},
    {
        "role": "user",
        "content": [
            {"type": "image", "image": "/path/to/image1"},
            {"type": "image", "image": "/path/to/image2"},
            {"type": "text", "text": "\nDescribe the two images in detail."},
        ],
    }
]

# Generate prompt and process inputs
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = processor(text=[prompt], images=images, return_tensors="pt").to(device)

# Generate output
generated_ids = model.generate(
    pixel_values=inputs.pixel_values,
    input_ids=inputs.input_ids,
    attention_mask=inputs.attention_mask,
    max_new_tokens=1024,
    do_sample=False,
    eos_token_id=tokenizer.eos_token_id,
)

# Decode output
output_text = processor.batch_decode(
    generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)[0]
print(output_text)

Video Inference (from Pre-extracted Frames)

Note: Video decoding is not included. You must first extract frames from your video using ffmpeg or similar tools.

Step 1: Extract frames from video

# Extract 1 frame per second
ffmpeg -i video.mp4 -vf fps=1 frames/frame_%04d.jpg

# Or extract a specific number of frames uniformly
ffmpeg -i video.mp4 -vf "select=not(mod(n\,30))" -vsync vfr frames/frame_%04d.jpg

Step 2: Run inference on frames

import torch
from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer

import video_io  # Helper module for loading frames

# Load model and processor
model_path = "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16"  # Or use a local path
device = "cuda:0"

model = AutoModelForCausalLM.from_pretrained(
    model_path,
    trust_remote_code=True,
    device_map=device,
    torch_dtype=torch.bfloat16
).eval()
tokenizer = AutoTokenizer.from_pretrained(model_path)
processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)

# Load frames from directory
frames_dir = "path/to/frames_directory"
video_fps = 1  # FPS used when extracting frames (for temporal understanding)

frames = video_io.load_frames_from_directory(frames_dir)
image_urls, metadata = video_io.frames_to_data_urls_with_metadata(frames, video_fps)

print(f"Loaded {len(frames)} frames, metadata: {metadata}")

# Prepare messages
messages = [
    {"role": "system", "content": "/no_think"},
    {
        "role": "user",
        "content": [
            {"type": "video", "video": ""},
            {"type": "text", "text": "\nDescribe what you see."},
        ],
    }
]

# Generate prompt and process inputs
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Process with FPS metadata for better temporal understanding
if metadata:
    inputs = processor(
        text=[prompt],
        videos=frames,
        videos_kwargs={'video_metadata': metadata},
        return_tensors="pt",
    )
else:
    inputs = processor(
        text=[prompt],
        videos=frames,
        return_tensors="pt",
    )
inputs = inputs.to(device)

# Set video pruning rate for efficient inference
model.video_pruning_rate = 0.75

# Generate output
generated_ids = model.generate(
    pixel_values_videos=inputs.pixel_values_videos,
    input_ids=inputs.input_ids,
    attention_mask=inputs.attention_mask,
    max_new_tokens=128,
)

# Decode output
output_text = processor.batch_decode(
    generated_ids, skip_special_tokens=False, clean_up_tokenization_spaces=False
)[0]
print(output_text)

Inference with vLLM

Make sure to use the main branch of vLLM. Run the following install command:

!VLLM_USE_PRECOMPILED=1 pip install git+https://github.com/vllm-project/vllm.git@main

Commands to serve the model:

BF16

vllm serve nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 --trust-remote-code --dtype bfloat16 --video-pruning-rate 0

FP8

vllm serve nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-FP8 --trust-remote-code --quantization modelopt --video-pruning-rate 0

FP4

vllm serve nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-NVFP4-QAD --trust-remote-code --quantization modelopt_fp4 --video-pruning-rate 0

Inference with SGLang

Support is verified in main:

pip install "git+https://github.com/sgl-project/sglang.git@main#subdirectory=python"

BF16

sglang serve --trust-remote-code --model-path nvidia/Nemotron-Nano-12B-v2-VL-BF16 --max-mamba-cache-size 256 # Adjust '--max-mamba-cache-size' as needed, to fit in memory

FP8

sglang serve --trust-remote-code --model-path nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-FP8

FP4

sglang serve --trust-remote-code --model-path nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-NVFP4-QAD --quantization modelopt_fp4

Training, Testing, and Evaluation Datasets:

Training Datasets:

Data Modalities
** Total Size: 39'486'703 samples
** Total Number of Datasets: 270

** Text-only datasets: 33
** Text-and-image datasets: 176
** Video-and-text datasets: 61
** Total size: 27.7 TB

** Data modalities: Text, Image, Video
** Data Collection Method by dataset: Hybrid: Automated, Human, Synthetic
** Labeling Method by dataset: Hybrid: Automated, Human, Synthetic

** Dataset partition: Training [100%], Testing [0%], Validation [0%]
** Time period for training data collection: 2023-2025
** Time period for testing data collection: N/A
** Time period for validation data collection: N/A

The post-training datasets consist of a mix of internal and public datasets designed for training vision language models across various tasks. It includes:

• Public datasets sourced from publicly available images and annotations, supporting tasks like classification, captioning, visual question answering, conversation modeling, document analysis and text/image reasoning.
• Internal text and image datasets built with public commercial images and internal labels, adapted for the same tasks as listed above.
• Synthetic image datasets generated programmatically for specific tasks like tabular data understanding and optical character recognition (OCR), for English, Chinese as well as other languages.
• Video datasets supporting video question answering and reasoning tasks from publicly available video sources, with either publicly available or internally generated annotations.
• Specialized datasets for safety alignment, function calling, and domain-specific tasks (e.g., science diagrams, financial question answering).
• NVIDIA-Sourced Synthetic Datasets for text reasoning.
• Private datasets for safety alignment or VQA on invoices.
• Crawled or scraped captioning, VQA, and video datasets.
• Some datasets were improved with Qwen2.5-72B-Instruct annotations

For around ~30% of our total training corpus and several of the domains listed above, we used commercially permissive models to perform:

• Language translation
• Re-labeling of annotations for text, image and video datasets
• Synthetic data generation
• Generating chain-of-thought (CoT) traces

Additional processing for several datasets included rule-based QA generation (e.g., with templates), expanding short answers into longer responses, as well as proper reformatting. More details can be found here.

** Image based datasets were all scanned against known CSAM to make sure no such content was included in training.

Public Datasets

Private Datasets

Data Crawling and Scraping

User-Sourced Data (Collected by Provider including Prompts)

Self-Sourced Synthetic Data

Properties

• Additionally, the dataset collection (for training and evaluation) consists of a mix of internal and public datasets designed for training and evaluation across various tasks. It includes:
  • Internal datasets built with public commercial images and internal labels, supporting tasks like conversation modeling and document analysis.
  • Public datasets sourced from publicly available images and annotations, adapted for tasks such as image captioning and visual question answering.
  • Synthetic datasets generated programmatically for specific tasks like tabular data understanding.
  • Specialized datasets for safety alignment, function calling, and domain-specific tasks (e.g., science diagrams, financial question answering).

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

Evaluation Datasets:

The following external benchmarks are used for evaluating the model:

Data Collection Method by dataset:

• Hybrid: Human, Automated
  

Labeling Method by dataset:

• Hybrid: Human, Automated
  

Properties (Quantity, Dataset Descriptions, Sensor(s)): N/A

Dataset License(s): N/A

Evaluation benchmarks scores:

Inference:

Acceleration Engine: vLLM
Acceleration Engine: TRT-LLM
Acceleration Engine: SGLang

Test Hardware:

• NVIDIA L40S
  
• NVIDIA A100
  
• NVIDIA B200
  
• NVIDIA H100/H200
  
• NVIDIA RTX PRO 6000 Server Edition
  
• NVIDIA GB200
  

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.

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

Please report security vulnerabilities or NVIDIA AI Concerns here.

Citation:

@misc{nvidia2025nvidianemotronnanov2,
      title={NVIDIA Nemotron Nano V2 VL},
      author={NVIDIA},
      year={2025},
      eprint={2511.03929},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2511.03929},
}