Run Multimodal Reasoning Agents with NVIDIA Nemotron on vLLM

We are excited to release NVIDIA Nemotron Nano 2 VL, supported by vLLM. This open vision language model (VLM) is built for video understanding and document intelligence.

Nemotron Nano 2 VL uses a hybrid Transformer–Mamba design and delivers higher throughput while maintaining state-of-the-art multimodal reasoning accuracy. The model also features Efficient Video Sampling (EVS), a new technique that reduces redundant tokens generation for video workloads , allowing processing of more videos with higher efficiency.

In this blog post, we’ll explore how Nemotron Nano 2 VL advances video understanding and document intelligence, showcase real-world use cases and benchmark results, and guide you through getting started with vLLM for inference to unlock high-efficiency multimodal AI at scale.

About Nemotron Nano 2 VL

  • Architecture:
    • CRADIOH-V2 based Vision Encoder
    • Efficient video sampling as token compression module
    • Hybrid Transformer-Mamba Architecture- Nemotron Nano 2 LLM backbone with reasoning.
  • Accuracy
    • Leading accuracy on OCRBench v2
    • 74 on average score (compared to 64.2 with current top VL model) on the following benchmarks: MMMU, MathVista, AI2D, OCRBench, OCRBench-v2, OCR-Reasoning, ChartQA, DocVQA, and Video-MME
  • Model size: 12B
  • Context length: 128k
  • Model input: Multi-image documents, videos, text
  • Model output: Text
  • Get started:
    • Download model weights from Hugging Face - BF16, FP8, FP4-QAD
    • Run with vLLM for inference
    • Technical report to build custom, optimized models with Nemotron techniques..

Run optimized inference with vLLM

Nemotron Nano 2 VL, achieves accelerated inference and serves more requests on the same GPU with BF16, FP8 and FP4 precision support. Follow these instructions to get started:

Create a fresh conda environment

```bash
conda create -n nemotron-vllm-env python=3.10 -y
conda activate nemotron-vllm-env
```

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

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

We can then serve this model via an OpenAI-compatible API


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

# FP8
```bash
vllm serve nvidia/Nemotron-Nano-VL-12B-V2-FP8 --trust-remote-code --quantization modelopt --video-pruning-rate 0
```

# FP4
```bash
vllm serve nvidia/Nemotron-Nano-VL-12B-V2-FP4-QAD --trust-remote-code --quantization modelopt_fp4 --video-pruning-rate 0
```

Once the server is up and running, you can prompt the model using the below code snippets

from openai import OpenAI
client = OpenAI(base_url="http://127.0.0.1:8033/v1", api_key="null")
# Simple chat completion
resp = client.chat.completions.create(
    model="nvidia/Nemotron-Nano-12B-v2-VL-BF16",
    messages=[
        {"role": "system", "content": "/no_think"},
        {"role": "user", "content": [
            {"type": "text", "text": "Give me 3 interesting facts about this image."}, 
            {"type": "image_url", "image_url": {"url": "https://blogs.nvidia.com/wp-content/uploads/2025/08/gamescom-g-assist-nv-blog-1280x680-1.jpg"}
            }
            ]},
    ],
    temperature=0.0,
    max_tokens=1024,
)
print(resp.choices[0].message.content)

For an easier setup with vLLM, refer to our getting started cookbook, available here.

Leading multimodal model for efficient video understanding and document intelligence

NVIDIA Nemotron Nano 2 VL brings both video understanding and document intelligence capabilities together in a single, highly efficient model. Built on the hybrid Transformer–Mamba architecture, it combines the reasoning strength of Transformer models with the compute efficiency of Mamba, achieving high throughput and low latency, allowing it to process multi-image inputs faster.

Trained on NVIDIA-curated, high-quality multimodal data, Nemotron Nano 2 VL leads in video understanding and document intelligence benchmarks such as MMMU, MathVista, AI2D, OCRBench, OCRBench-v2, OCR-Reasoning, ChartQA, DocVQA, and Video-MME, delivering top-tier accuracy in multimodal reasoning, character recognition, chart reasoning, and visual question answering. This makes it ideal for building multimodal applications that automate data extraction and comprehension across videos, documents, forms, and charts with enterprise-grade precision.


Figure 1: Nemotron Nano 2 VL provides leading accuracy on various video understanding and document intelligence benchmarks

Improving Efficiency with EVS
With EVS, the model achieves higher throughput and faster response times without sacrificing accuracy. EVS technique prunes redundant frames, preserving semantic richness while enabling longer video processing efficiently. As a result, enterprises can analyze hours of footage, from meetings and training sessions to customer calls, in minutes, gaining actionable insights faster and at lower cost.


Figure 2: Accuracy trend of the Nemotron Nano 2 VL model across various token-drop thresholds using efficient video sampling on Video-MME and LongVideo benchmarks

Get Started

To summarize, Nemotron Nano 2 VL helps build scalable, cost-efficient agentic AI systems that truly understand documents and video. With open weights, training datasets, and recipes, developers gain full transparency and flexibility to fine-tune and deploy the model across any environment, from on-premise to cloud, for maximum security and privacy.

Ready to build enterprise-ready agents?

Share your ideas and vote on what matters to help shape the future of Nemotron.

Stay up to date on NVIDIA Nemotron by subscribing to NVIDIA news and following NVIDIA AI on LinkedIn, X, YouTube, and the Nemotron channel on Discord.