Nemotron 3.5 Super VL and Bolt: NVIDIA's new open models you can download today
NVIDIA released Nemotron 3.5 Super VL, an open 120B multimodal agent model, plus Bolt 74B-A9B, which fits in 24GB for local use. Both are out now. A smaller 4GB Bolt E4B arrives October 23.
On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter vision-language model built for AI agents, together with Nemotron 3.5 Bolt 74B-A9B, a local model that fits in 24GB of memory. Both can be downloaded now. A third model, Bolt E4B (4GB), is due on October 23.
What NVIDIA released in Nemotron 3.5
The announcement covers three open models, each aimed at a different size of deployment:
- Nemotron 3.5 Super VL: a 120B-parameter multimodal model (the Hugging Face listing is named 120B-A12B) for enterprise agents that run all the time. It reads images, web pages and PDFs, including OCR, tables and charts, and is designed for tool use, following instructions and business workflows with many steps.
- Nemotron 3.5 Bolt 74B-A9B 24GB: a model for everyday agent tasks on local hardware. NVIDIA uses native 2-bit quantization so it fits in 24GB and can run next to other applications.
- Nemotron 3.5 Bolt E4B 4GB: a small model for voice assistants, interactive characters and embedded interfaces. It is coming October 23.
NVIDIA says Super VL matches the accuracy of its biggest reasoning model on multimodal agent tasks while using about a fifth of the parameters. The company also claims it is five times faster than the next-best open agentic model in its class, citing the Artificial Analysis leaderboard. For Bolt 74B-A9B, NVIDIA says it ranks first among open models of its class on PinchBench and second on BrowseComp. All of these are NVIDIA's own claims, and the post includes no independent testing.
How Super VL and Bolt fit the Nemotron lineup
The new models join two earlier members of the family. Lightning, released in August, handles high-throughput workloads. Ultra covers the hardest reasoning tasks. Super VL sits in the middle for multimodal agents, and Bolt brings the family to edge devices and local machines.
NVIDIA is presenting openness as the main difference from closed frontier models. Super VL comes with open weights, open datasets and training recipes, so organizations can post-train it on their own data and run it on their own infrastructure. According to the post, the model was trained on about 26 trillion tokens, then post-trained on roughly 3.8 million samples with reinforcement learning across dozens of environments. NVIDIA is also publishing datasets on Hugging Face covering tool use, finance, instruction following, long context, fairness and vision-language tasks, along with reinforcement learning environments in NeMo Gym.
The company names several software firms that are integrating Super VL, including Atlassian (agentic search across Jira and Confluence), CrowdStrike (security blue teaming), Perplexity (search and agentic tasks) and Salesforce (government workflow agents).
Who should care about Nemotron 3.5
- Developers building agents who want an open multimodal model they can host themselves. Super VL is aimed at document-heavy work, such as pulling data from PDFs, charts and web pages.
- Teams with data-residency or privacy requirements. Because the weights are open, the model can run on premises, so data never has to leave the company's own systems.
- Local AI users with a 24GB GPU or an NVIDIA Spark machine. Bolt 74B-A9B targets this group, and NVIDIA says it can handle web research, writing and code debugging through local agent harnesses without sending any work to the cloud.
- Hardware buyers. Bolt is built for the NVIDIA RTX Spark laptops that NVIDIA says arrive on October 16, and it also runs on GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations.
If you only use hosted chatbots and never run models yourself, nothing changes for you right now.
Where to download and run the models
NVIDIA lists these options:
- Super VL: open weights on Hugging Face, plus access through OpenRouter, the ModelScope community and as an NVIDIA NIM microservice. NVIDIA says the NIM version is updated every two weeks with security and performance fixes. ModelScope supports fine-tuning through its ms-swift framework.
- Bolt 74B-A9B 24GB: available on Hugging Face, with support in llama.cpp, Ollama, LM Studio, Unsloth and vLLM. A GGUF build is listed for Ollama.
- Customization: through NVIDIA NeMo, Python pipelines and the open datasets.
For large-scale serving, NVIDIA says that running Super VL on its Dynamo framework lets organizations serve up to eight times more users with 85% lower latency. This is another vendor figure. Super VL runs on DGX Spark, DGX Station, RTX PRO and RTX systems, Jetson edge devices, and data center and cloud environments.
Frequently asked questions
Is Nemotron 3.5 Super VL open source?
NVIDIA releases it with open weights, open datasets and training recipes, so you can download, post-train and host it yourself.
Can I run Nemotron 3.5 Bolt on my own PC?
Bolt 74B-A9B is designed to fit in 24GB of memory using 2-bit quantization. It runs through llama.cpp, Ollama, LM Studio, Unsloth or vLLM on GeForce RTX PCs and NVIDIA Spark systems.
When is Bolt E4B available?
NVIDIA says the 4GB Bolt E4B model arrives on October 23, 2026. Super VL and Bolt 74B-A9B are available now.
What can Super VL read?
It handles images, web pages and PDFs, including OCR, tables and charts, so an agent can pull information from documents without sending them to a separate vision system.
