NVIDIA has released Nemotron 3.5 Super VL, an open 120-billion-parameter vision-language model built for specialized, always-on AI agents, alongside two smaller Bolt models designed to run locally on edge hardware. Why it matters: NVIDIA is giving enterprises and developers a powerful open alternative to closed frontier models — one they can post-train on their own data and run on their own infrastructure.
October 15, 2026 — The new models were released on Thursday, October 15.
What NVIDIA released
Super VL is the flagship. It handles text, images, documents and video in a single model, so an agent can read a chart, a scanned page and a written report without a separate vision model alongside it. NVIDIA positions it for long-running agent work such as coding, document processing, research and video review.
Two smaller models join it under the Bolt name. Bolt 74B-A9B is available now and is squeezed into 24GB of memory, aimed at local agents on RTX laptops and workstations. Bolt E4B fits in 4GB, targets fast conversational and embedded agents, and is scheduled for October 23, 2026.
The details that matter
Super VL uses a mixture-of-experts design: it holds 120 billion parameters in total but activates only about 12 billion per token, which is how NVIDIA keeps it comparatively fast. Its model card on Hugging Face lists support for context lengths up to 1 million tokens, a Mamba-2 and Transformer hybrid design, and inputs spanning text, images and short video. It is released for commercial and non-commercial use under the OpenMDW license.
NVIDIA says the model delivers up to five times higher throughput than other open models in its size class on a single GPU, with accuracy it describes as comparable to its largest reasoning model at roughly a fifth of the parameter count. Treat those as the company’s own reported figures for now — independent verification is still pending.
Availability is broad from day one. Super VL is on Hugging Face, OpenRouter and ModelScope, and is also offered as an NVIDIA NIM microservice. The Bolt 74B-A9B model supports common local runtimes including llama.cpp, Ollama, LM Studio, Unsloth and vLLM, and the timing lines up with NVIDIA RTX Spark laptops arriving October 16.
Who is already building on it
According to NVIDIA, dozens of organizations have post-trained Super VL on their own data, including Scale AI working on finance agents and AfterQuery on credit-agreement analysis. Platform companies named as integrating the model include Atlassian, CrowdStrike, Salesforce and Perplexity. Those are NVIDIA’s reported adopters, not independently confirmed deployments.
Why it matters
Open-weight models keep closing the gap that once made closed APIs the default choice. A 120B model that activates only 12B parameters per token, reads documents and video, and can be fine-tuned and self-hosted gives businesses a real option for agents that handle sensitive data — the workload NVIDIA’s hardware business most wants to serve. For developers, the practical test starts now: download it, run it on your own tasks, and compare the results against what you pay for today. That hands-on comparison, not the launch benchmarks, will decide whether Super VL earns a place in real agent stacks.
Frequently asked questions
Is Nemotron 3.5 Super VL free to use?
The model is open and available on Hugging Face, OpenRouter and ModelScope for commercial and non-commercial use under the OpenMDW license. Running it still requires suitable GPU hardware or a hosting provider.
How big is Nemotron 3.5 Super VL?
It has 120 billion total parameters and activates about 12 billion per token. Its context window supports up to 1 million tokens.
What are the Bolt models for?
Bolt models are smaller versions for local and edge hardware. The 74B-A9B version fits in 24GB of memory and is available now; the 4GB E4B version is scheduled for October 23, 2026.
Can Super VL understand images and documents?
Yes. It is a vision-language model that takes text, images and video as input, including charts, tables and scanned documents.
Sources: NVIDIA (Hugging Face model card), Unite.AI