# Nvidia Nemotron Nano 12B V2 VL

A 12B vision-language model for document intelligence and video understanding. Nvidia Nemotron Nano 12B V2 VL uses a hybrid Mamba-Transformer architecture with document OCR and efficient video sampling.

- **Model ID:** `nvidia/nemotron-nano-12b-v2-vl`
- **Type:** chat
- **Providers:** bedrock
- **Context window:** 131,072
- **Maximum output tokens:** 131,072
- **Pricing:** $0.2/1M input tokens, $0.6/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/nemotron-nano-12b-v2-vl

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'nvidia/nemotron-nano-12b-v2-vl',
  prompt: 'Why is the sky blue?'
})
```

## About

Nvidia Nemotron Nano 12B V2 VL is an open 12B multimodal reasoning model from NVIDIA, released on October 28, 2025. It handles document intelligence and video understanding tasks, letting agents extract, interpret, and act on information across text, images, tables, and videos in a single model. At launch, Nvidia Nemotron Nano 12B V2 VL cited OCRBenchV2, reflecting document-level optical character recognition (OCR) and structured extraction capability.

The architecture is a hybrid Mamba-Transformer, the same design philosophy as the broader Nemotron family but applied to vision-language tasks. For video inputs, the model implements Efficient Video Sampling (EVS). EVS identifies and prunes temporally static patches, reducing token redundancy and processing longer video clips with up to 2.5x higher throughput without accuracy loss.

Nvidia Nemotron Nano 12B V2 VL runs on vLLM and TRT-LLM inference engines. Embedding and retrieval models in the same family appear on leaderboards such as ViDoRe, MTEB, and MMTEB for visual, multimodal, and multilingual text retrieval. The NVIDIA AI Blueprint for video search and summarization (VSS) is built around this model, making it a practical foundation for production video intelligence pipelines. Announcement and agent-focused context: https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html.

## What to consider

Video inputs consume significantly more tokens than static images. Estimate average clip length and token density before deploying video tasks at scale so you can budget token costs accurately. Compare $0.2 and $0.6.

## When to use

### Best For

- Document intelligence: extracting structured text, tables, and bounding boxes from PDFs, scanned forms, and reports
- Video understanding: dense captioning, video Q&A, or summarization of longer clips with efficient token handling
- Multimodal retrieval-augmented generation (RAG) pipelines that retrieve and reason across documents with images, diagrams, and tables alongside text
- Media asset management workflows requiring semantic search and retrieval across visual content
- Agentic systems that perceive and act on mixed-modality inputs in a single model call

### Consider Alternatives When

- Your task is text-only. Nemotron 3 Nano or Super are better fits
- You need maximum reasoning depth for complex text-based multi-agent planning. Super's architecture is optimized for that role
- Your video inputs are short and simple, and a lighter vision model would suffice
- You need latent MoE or multi-token prediction for throughput-optimized long-context generation

## Best for

- **Document intelligence:** Extracting structured text, tables, and bounding boxes from PDFs, scanned forms, and reports
- **Video understanding:** Dense captioning, video Q&A, or summarization of longer clips with efficient token handling
- **Multimodal RAG pipelines:** Retrieval and reasoning across documents with images, diagrams, and tables alongside text
- **Media asset management:** Workflows requiring semantic search and retrieval across visual content
- **Mixed-modality agents:** Agentic systems that perceive and act on text, image, and video inputs in a single model call

## Consider alternatives

- **Text-only workloads:** Nemotron 3 Nano or Super are better fits without the multimodal overhead
- **Deep text reasoning:** Super's architecture is optimized for complex text-based multi-agent planning
- **Short simple videos:** A lighter vision model may suffice for short video inputs
- **Throughput-optimized generation:** Models with latent MoE or multi-token prediction handle long-context generation more efficiently

## Frequently asked questions

### What types of image inputs does Nvidia Nemotron Nano 12B V2 VL support?

The model handles image Q&A, OCR, dense captioning, and multi-image reasoning. Nvidia Nemotron Nano 12B V2 VL cited OCRBenchV2 at launch. OCRBenchV2 tests text extraction from document images with complex layouts, tables, and mixed formatting.

### What is Efficient Video Sampling (EVS)?

EVS identifies and prunes temporally static patches in video sequences (frames where little changes between consecutive images). Removing redundant patches reduces the token count per video clip. The model can process longer videos with up to 2.5x higher throughput without sacrificing accuracy.

### How does this model support RAG pipelines?

Nvidia Nemotron Nano 12B V2 VL serves as the reasoning component for visual content in the Nemotron RAG suite. Embedding models in the same family appear on ViDoRe, MTEB, and MMTEB leaderboards for visual, multimodal, and multilingual text retrieval. Together, they enable retrieval-augmented generation (RAG) across proprietary data with mixed-modality documents.

### What benchmark did Nvidia Nemotron Nano 12B V2 VL highlight at launch?

OCRBenchV2. It measures document intelligence and optical character recognition on visually complex documents.

### Is this model open source?

Yes. NVIDIA released model weights on Hugging Face under the NVIDIA Open Model License.

### Can I use this model for multi-image reasoning tasks?

Yes. Multi-image reasoning is part of the model's task coverage across image Q&A, OCR, dense captioning, video Q&A, and multi-image reasoning. You can use it for tasks like comparing document versions, analyzing image sequences, or reasoning over slide decks.

### Where are per-token prices listed?

Rates are listed on this page. They reflect the providers routing through AI Gateway and shift when providers update their pricing.

## Links

- [Model page](https://vercel.com/ai-gateway/models/nemotron-nano-12b-v2-vl)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://docs.aws.amazon.com/bedrock/latest/userguide/models-supported.html)
- [Provider pricing](https://aws.amazon.com/bedrock/pricing/)