Key Takeaways
- NVIDIA Nemotron 3.5 Super VL is an open vision-language model from NVIDIA AI, announced October 15, 2026, with 120B total parameters and 12B active per request.
- It handles text, images, documents and video in one model, and is designed for long-running AI agents that need reasoning and tool use across many steps.
- No dedicated product page exists yet; check NVIDIA's official site at build.nvidia.com for availability, license terms and access details.
- The 120B/12B design keeps large model capacity while limiting compute per request, similar to other sparse or mixture-of-experts architectures.
Quick Facts
| Type | Open vision-language model for AI agents |
|---|---|
| Devices | Server/cloud deployment; check official site for details |
| Price | Not announced; check official site |
| Languages | Not specified; likely multilingual, check official site |
| Highlight | 120B total, 12B active; text, image, document, video |
| Learning curve | Moderate for developers familiar with LLM APIs |
| Category | AI & Models |
| Creator | NVIDIA AI |
| Official site | build.nvidia.com ↗ |
| Heat | 🔥 602.1K Heat |
| Last updated |
What NVIDIA Nemotron 3.5 Super VL Actually Is
NVIDIA Nemotron 3.5 Super VL is the newest member of NVIDIA's Nemotron family of open models, announced on October 15, 2026. The "VL" stands for vision-language, meaning it understands not just text but also images, documents and video. NVIDIA positions it as a single model that combines reasoning, tool use and multimodal understanding, specifically built for AI agents that keep working on a task for a long time. This is a notable shift from using separate specialist models for each input type.
The model has 120 billion total parameters but only 12 billion active at any one time. This large-total, small-active design is generally used to keep plenty of model capacity while holding down the compute each request needs. It is similar in spirit to mixture-of-experts architectures, where only part of the network fires for a given input. The result is a model that can be powerful yet more efficient to run than a dense 120B model would be.
NVIDIA has not yet published a dedicated page for this model. The announcement did not list a context length, benchmark scores or license terms, so those details remain unconfirmed. For now, the official site at build.nvidia.com is the place to watch for updates. The word "open" in the announcement is significant: it suggests teams will be able to host the model themselves, which matters for privacy, cost control and customization.
- Open weights for self-hosting
The model is described as open, meaning teams may be able to download and run it on their own infrastructure. This is a major advantage for organizations with strict data privacy needs.
- Multimodal by design
It reads text, images, documents and video in one model, so an agent can process a screenshot, parse a PDF and follow a screen recording without switching between specialist models.
- Built for long-running agents
The combination of reasoning and tool use targets agents that work through multi-step jobs, such as researching a topic, calling APIs and summarizing findings over many turns.
- Efficient active parameter count
With 12B active out of 120B total, the model aims to balance capability and compute cost, making it more practical for sustained agent workloads.
How the Vision-Language and Agent Capabilities Work
At its core, Nemotron 3.5 Super VL is designed to handle the full loop of an agentic task. An agent working on a multi-step job often has to read a screenshot, parse a PDF, follow a screen recording and then call a tool. This model is built to handle that entire loop in a single model rather than a chain of specialist ones. That reduces the complexity of orchestrating multiple models and can improve reliability because the same model maintains context across different input types.
The vision-language side means the model can interpret visual information alongside text. For example, it could look at a chart in an image, extract the data, and then reason about it in text. It could also read a document with mixed text and diagrams, or analyze a video to understand a process. This is particularly useful for agents that need to interact with real-world interfaces, such as browsing a website or operating software.
The reasoning and tool use components allow the model to plan and execute actions. Instead of just answering a question, it can decide to call an external tool, such as a search engine or a calculator, and then incorporate the result into its response. This makes it suitable for tasks that require multiple steps and external data. The 12B active parameter design helps keep inference fast enough for interactive agent loops, though exact latency will depend on hardware and deployment.
- Unified multimodal understanding
Instead of stitching together a vision model and a language model, Nemotron 3.5 Super VL handles text, images, documents and video natively, which can reduce errors at the seams.
- Tool use for real actions
The model can call external tools as part of its reasoning, enabling agents to fetch live data, run code or interact with APIs without human intervention.
- Long-horizon task support
It is optimized for agents that run for many steps, maintaining context and goal direction over extended interactions, which is harder for smaller or less reasoning-focused models.
- Sparse activation for efficiency
Only 12B of the 120B parameters are active per request, which helps control compute costs while preserving a large knowledge base.
How to Access and Use Nemotron 3.5 Super VL
As of the announcement, there is no dedicated product page for Nemotron 3.5 Super VL. NVIDIA's official site at build.nvidia.com is the primary place to check for updates. Since the model is described as open, it is likely that weights will be made available for download, possibly through NVIDIA's NGC catalog or Hugging Face. However, exact access methods, licensing and pricing have not been confirmed. Developers should monitor the official site for the latest information.
If you are planning to use the model, you will need infrastructure capable of running a 120B-parameter model, even with only 12B active. That typically means a server with sufficient GPU memory, such as NVIDIA's own data center GPUs. For teams that want to host their own models and still need vision, the word "open" is the real headline here. Self-hosting gives you control over data, customization and cost, but requires technical expertise.
Once available, using the model will likely involve an API or a local inference framework. You would send it a prompt that can include text, images, documents or video, and it would return a response that may include tool calls. For agentic workflows, you would integrate it with a tool-calling framework that lets the model invoke external functions. The exact API shape and supported formats are not yet published, so check the official documentation when it appears.
- Watch the official site
build.nvidia.com is where NVIDIA will likely post model cards, download links and documentation. Check there first for availability.
- Prepare your infrastructure
Running a 120B model requires significant GPU resources. Plan for data center-class hardware or a cloud instance with enough VRAM.
- Familiarize yourself with agent frameworks
Since the model is built for tool use, learning a framework like LangChain or NVIDIA's own agent tools will help you integrate it effectively.
- Test with multimodal inputs
Once you have access, try feeding it a mix of text, images and PDFs to see how well it maintains context across modalities in a single session.
How It Compares and Who It Is For
Nemotron 3.5 Super VL enters a growing field of vision-language models. Compared to closed models like GPT-4V or Gemini, its main differentiator is being open, which allows self-hosting and customization. Compared to other open VLMs, its 120B total parameter count is on the larger side, but the 12B active design may make it more efficient than dense models of similar capability. However, without benchmark scores, it is hard to say how it stacks up in accuracy or speed.
The model is clearly aimed at developers and organizations building AI agents that need to process multimodal data over long tasks. This includes use cases like automated customer support that reads screenshots, research assistants that analyze documents and video, and robotic process automation that interacts with graphical interfaces. It is less suited for simple, single-turn question answering where a smaller model would suffice.
Limitations and open questions remain. The lack of published context length, license terms and benchmarks means potential users cannot yet evaluate fit. The compute requirements for self-hosting may be prohibitive for smaller teams. Also, the model's performance on specific tasks like video understanding or document parsing is unproven. As with any new model, early adopters should test thoroughly and watch for updates from NVIDIA.
- Open vs closed
Unlike proprietary models, Nemotron 3.5 Super VL can potentially be self-hosted, giving teams more control over data and costs.
- Efficiency vs dense models
The 12B active parameter design may offer better throughput than a dense 120B model, but actual performance depends on the task and hardware.
- Ideal for agentic workflows
It is best suited for multi-step tasks that combine vision, language and tool use, rather than simple text generation.
- Unconfirmed details
License, context length, supported languages and benchmarks are not yet announced. Check the official site for updates before committing.
Step-by-Step Guide
Check the official site for availability
Go to build.nvidia.com and look for Nemotron 3.5 Super VL. Since there is no dedicated page yet, you may need to search the site or check NVIDIA's developer blog for announcements. Sign up for updates if possible.
Review the model card and license
Once available, read the model card carefully. It should list context length, supported input formats, license terms and any usage restrictions. This will tell you if the model fits your project and compliance needs.
Prepare your environment
Decide whether you will run the model locally or via a cloud API. For local deployment, ensure you have GPUs with enough memory. For API access, get the necessary credentials and review rate limits and pricing.
Integrate with your agent framework
Use an agent framework that supports tool calling and multimodal inputs. Connect the model to your tools and test with simple tasks first, such as extracting text from an image or summarizing a PDF.
Test and iterate on real tasks
Start with a representative task, like having the agent read a screenshot and call an API. Monitor performance, accuracy and cost. Adjust prompts and tool definitions as needed, and stay tuned for official updates.
Frequently Asked Questions
What is NVIDIA Nemotron 3.5 Super VL?
It is an open vision-language model from NVIDIA AI, announced on October 15, 2026. It combines reasoning, tool use and multimodal understanding for text, images, documents and video. With 120B total parameters and 12B active, it is designed for long-running AI agents that need to process different types of data in one model.
How can I access Nemotron 3.5 Super VL?
There is no dedicated product page yet. Check NVIDIA's official site at build.nvidia.com for updates. Since the model is described as open, it may be released for download or via an API. Exact access methods, pricing and licensing have not been announced, so monitor the official site for details.
What does 120B total and 12B active mean?
It means the model has 120 billion parameters in total, but only 12 billion are used for each request. This sparse activation, similar to mixture-of-experts, allows the model to have a large capacity while keeping compute costs lower. It can make inference faster and cheaper than a dense 120B model.
What can I do with Nemotron 3.5 Super VL?
You can build AI agents that read screenshots, parse PDFs, analyze video and call external tools. It is suited for multi-step tasks like automated research, customer support that handles visual data, or robotic process automation. It is less useful for simple text-only tasks where a smaller model would suffice.
How does it compare to other vision-language models?
Its main advantage is being open, which allows self-hosting and customization, unlike closed models such as GPT-4V. Its 120B/12B design may offer efficiency gains over dense models. However, without published benchmarks, it is hard to compare accuracy. Check the official site for performance data when available.
