README.md

---
library_name: transformers
license: other
license_name: nscl-v1
pipeline_tag: image-text-to-text
tags:
- nvidia
- pytorch
- multimodal
- vlm
- diffusion-language-model
---

# Nemotron-Labs-Diffusion-VLM-8B


<div align="center" style="line-height: 1;">
<a href="https://d1qx31qr3h6wln.cloudfront.net/publications/Nemotron_Diffusion_Tech_Report_v1.pdf?VersionId=db8_EMO8B.vmU26.jr7Le9pN3MqcUDNL" target="_blank" style="margin: 2px;">
    <img alt="Chat" src="https://img.shields.io/badge/📝Paper-Read Now!-536af5?color=76B900&logoColor=white" style="display: inline-block; vertical-align: middle;"/>
</a>
<a href="https://huggingface.co/collections/nvidia/nemotron-labs-diffusion" target="_blank" style="margin: 2px;">
    <img alt="Nemotron-Labs-Diffusion Model Family" src="https://img.shields.io/badge/%F0%9F%A4%97-Nemotron--Labs--Diffusion_Model_Family-76B900" style="display: inline-block; vertical-align: middle;"/>
</a>
<a href="https://www.nvidia.com/en-us/agreements/enterprise-software/nvidia-source-code-license/" style="margin: 2px;">
  <img alt="License" src="https://img.shields.io/badge/License-NSCLv1-f5de53?&color=f5de53" style="display: inline-block; vertical-align: middle;"/>
</a>
</div>


[![Demo](./assets/demo.gif)](./assets/demo.mp4)


## Model Overview

Nemotron-Labs-Diffusion-VLM-8B is the vision-language extension of the Nemotron-Labs-Diffusion family. It pairs the same tri-mode language backbone (AR / diffusion / self-speculation, switchable by attention pattern) with a vision encoder, accepting interleaved image + text input and producing text output. The diffusion-based parallel decoding from the LM family carries over to VLM: the language head can draft a block in parallel and verify autoregressively against shared KV cache, retaining the family's decode-efficiency story while extending it to multimodal prompts.

<div align="center">
<img src="./assets/teaser.png" alt="An illustration of Tri-Mode LMs" width="500">
</div>


## Key Design

- 8B vision-language model in the Nemotron-Labs-Diffusion family — same tri-mode language backbone (AR, diffusion, self-speculation) plus a Pixtral-style vision encoder.
- Vision encoder: 24-layer, 1024-hidden, 14×14 patch, supports up to 1540×1540 images with `spatial_merge_size=2`.
- Language decoder weights match `nvidia/Nemotron-Labs-Diffusion-8B` (34 layers, 4096 hidden, 14336 intermediate); the model card structure and inference modes inherit from the LM line.
- Diffusion-based parallel decoding works for multimodal prompts: image tokens are placed in the bidirectional context window and text generation proceeds via the same block-wise unmasking + AR verification as the LM family.


## License/Terms of Use

Use of this model is governed by the **NVIDIA Source Code License (NSCLv1)**.


## Environment

```bash
transformers>=5.0.0
pillow
requests
opencv-python
```


## Chat with Our Model

```python
import sys
import torch
from huggingface_hub import snapshot_download
from transformers import AutoModel, AutoTokenizer

repo_name = "nvidia/Nemotron-Labs-Diffusion-VLM-8B"
sys.path.insert(0, snapshot_download(repo_name))
from image_processing import process_messages

tokenizer = AutoTokenizer.from_pretrained(repo_name, trust_remote_code=True)
model = AutoModel.from_pretrained(repo_name, trust_remote_code=True).cuda().to(torch.bfloat16)

image_path = "path/to/your/image.jpg"  # local file or http(s):// URL
messages = [{
    "role": "user",
    "content": [
        {"type": "image_url", "image_url": {"url": image_path}},
        {"type": "text", "text": "Describe this image."},
    ],
}]

batch = process_messages(tokenizer, messages, add_generation_prompt=True)
prompt_ids = batch["input_ids"].to("cuda")
pixel_values = batch["pixel_values"].to("cuda", dtype=torch.bfloat16)

out_ids, nfe = model.generate(
    prompt_ids,
    pixel_values=pixel_values,
    image_sizes=batch["image_sizes"],
    max_new_tokens=512, steps=512, block_length=32,
    shift_logits=False, threshold=0.9,
    eos_token_id=tokenizer.eos_token_id,
)

tokenized_out = tokenizer.batch_decode(out_ids[:, prompt_ids.shape[1]:], skip_special_tokens=True)
print(f"Model: {tokenized_out[0]}")
print(f"[Num Function Eval (NFE)={nfe}]")
```


## 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 [bias](./model_cards/bias.md), [explainability](./model_cards/explainability.md), [safety & security](./model_cards/safety.md), and [privacy](./model_cards/privacy.md) subcards.

Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns [here](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).


## Citations

```bibtex
@techreport{fu2026nemotronlabsdiffusion,
  title       = {Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding},
  author      = {Yonggan Fu and Lexington Whalen and Abhinav Garg and Chengyue Wu and Maksim Khadkevich and Nicolai Oswald and Enze Xie and Daniel Egert and Sharath Turuvekere Sreenivas and Shizhe Diao and Chenhan Yu and Ye Yu and Weijia Chen and Sajad Norouzi and Shiyi Lan and Ligeng Zhu and Jin Wang and Jindong Jiang and Morteza Mardani and Mehran Maghoumi and Song Han and Ante Jukic and Nima Tajbakhsh and Jan Kautz and Pavlo Molchanov},
  institution = {NVIDIA},
  year        = {2026},
  note        = {Technical report}
}
```