NVIDIA Nemotron Nano 12B v2 VL
NVIDIA Nemotron Nano 12B v2 VL is an openly released ~12B-parameter vision-language model from NVIDIA's Nemotron Nano family, aimed at multimodal chat and document understanding at small scale.
API Pricing
| Provider | Input / 1M | Output / 1M |
|---|---|---|
| $0.200 | $0.600 |
Prices updated daily. Last check: Oct 12, 2026
Compare API pricing for every Nemotron model →NVIDIA Nemotron Nano 12B v2 VL pricing by provider
Input, output, and batch rates, plus alternatives, for one provider at a time.
Performance & Benchmarks
Source: Artificial Analysis →Reasoning & Knowledge
- MMLU-Pro64.9%
- GPQA Diamond43.9%
- Humanity's Last Exam4.3%
Coding
- LiveCodeBench34.5%
Math
- AIME 202526.7%
Agentic & Tool Use
- Terminal-Bench Hard0.0%
- τ²-bench19.3%
Instruction & Long Context
- IFBench25.9%
- Long-Context Reasoning20.7%
Benchmarks measured Oct 2026. Scores are independent evaluations, not vendor-reported.
Model Details
General
- Creator
- NVIDIA
- Modalities
- Text
Capabilities
- Open Source
- Yes
Strengths & Limitations
Strengths
- Open-source weights, so the model can be self-hosted on rented or owned GPUs instead of being locked to a single vendor API
- Vision-language configuration (the "VL" variant) handles prompts that combine images with text instructions
- Roughly 12B parameters places it in a size class that is often servable on a single modern accelerator, depending on precision and context length
- Part of NVIDIA's Nemotron Nano line, which targets the compact end of the family rather than maximum-scale deployments
- Open availability means multiple inference providers can host the same checkpoint, creating price competition visible in the table on this page
- Second-generation (v2) release within the Nemotron Nano series rather than a first iteration
Limitations
- We do not have a confirmed context window for this model in our database — verify the limit with your chosen provider
- Published throughput and time-to-first-token figures are not populated in our benchmark data, so latency must be measured per provider
- At roughly 12B parameters it is a compact model and is not positioned against large frontier-scale multimodal systems
- Hosted availability for Nemotron models is narrower than for the most widely served open models, which can limit provider choice
- Self-hosting shifts cost and operational burden to GPU provisioning, quantization, and serving-stack tuning
Key Features
About NVIDIA Nemotron Nano 12B v2 VL
Common Use Cases
Nemotron Nano 12B v2 VL suits workloads that pair an image with a text instruction at volume: document and form understanding, screenshot or UI description, chart and diagram question answering, image captioning and tagging pipelines, and visual content triage. Its compact size makes it a reasonable fit for batch processing jobs where thousands of images must be described or classified and where per-item cost matters more than ceiling capability on hard reasoning tasks. Because the weights are open, it is also a candidate for deployments with data-residency or air-gap requirements that rule out third-party APIs, and for teams that want to fine-tune a multimodal base on domain-specific imagery. For complex multi-step agentic reasoning or long-document analysis, evaluate larger models in the Nemotron line or hosted frontier-tier multimodal models instead, and confirm the context window your workload needs with the provider you choose.
Frequently Asked Questions
How much does NVIDIA Nemotron Nano 12B v2 VL cost to run?
Pricing depends on which provider you use and how you consume the model. Hosted endpoints bill per input and output token, and rates differ between providers serving the same open checkpoint. Because the weights are open, you can also self-host on rented GPUs and pay by the hour instead, which changes the economics entirely based on your utilization. Check the pricing table on this page for current per-provider rates, and compare against the GPU rental listings on this site if you are considering self-hosting.
What is Nemotron Nano 12B v2 VL best used for?
It is best suited to high-volume multimodal tasks — image captioning, document and form extraction, chart or screenshot question answering, and visual classification — where a compact vision-language model is adequate and per-item cost matters. It is also a practical choice when open weights are a requirement, such as on-premises or air-gapped deployments and domain-specific fine-tuning.
Is Nemotron Nano 12B v2 VL open source?
Yes. Our catalog records this model as open source, meaning NVIDIA has published the weights rather than restricting access to a proprietary API. Check the specific license terms NVIDIA attaches to the release before using it commercially, as open-weight releases vary in what they permit.
What context window does it support?
We do not have a confirmed context window for this entry in our database. Context limits for open models can also differ between hosted providers, since each serving deployment configures its own maximum. Verify the supported context length with the provider listed in the pricing table or with NVIDIA's model card before building against a specific limit.
How does the VL version differ from other Nemotron Nano models?
The "VL" suffix indicates a vision-language configuration, so this variant is intended to accept image input alongside text, whereas other Nemotron Nano entries in the family are text-focused. The "v2" marks it as a second-generation release in the Nano series, which sits at the compact end of NVIDIA's Nemotron lineup.
Should I use a hosted endpoint or self-host this model?
That depends on your volume and constraints. Hosted endpoints avoid infrastructure work and suit bursty or low-volume traffic, while self-hosting a ~12B model on rented GPU capacity can be more economical at sustained high throughput and is necessary if your data cannot leave your environment. Compare the per-token rates in the table on this page against the GPU hourly rates listed elsewhere on this site using your own expected request volume.