NVIDIA releases open Nemotron models that run on a single gaming GPU
NVIDIA has released Nemotron 3.5 Super VL, which reads documents and web pages, and the local Bolt models as open models. Bolt 74B-A9B fits in 24 GB of memory with 2-bit quantization, and the training data is open too.
On October 15, 2026, NVIDIA announced the Nemotron 3.5 family. It includes Super VL for enterprise agent work and the Bolt models for personal computers. Training datasets and recipes were released alongside the models, which can be downloaded from Hugging Face and run on your own servers or PCs.
Super VL reads documents and acts on them
Nemotron 3.5 Super VL is a 120-billion-parameter vision-language model. It reads images, web pages and PDFs, and NVIDIA says it is especially strong at text recognition, tables, charts and structured information in documents. It is built for agents that act on what they read: multi-step workflows, document processing, coding agents, insurance fraud scanning and credit review are among the examples. The model was pretrained on about 26 trillion tokens, then post-trained on about 3.8 million samples plus reinforcement learning across dozens of environments. It is available on Hugging Face, OpenRouter, ModelScope and NVIDIA NIM, and runs on hardware from DGX Spark to RTX workstations and Jetson.
Bolt models that run locally
Nemotron 3.5 Bolt 74B-A9B is built for local agent work. Of its 74 billion parameters, 9 billion are active at each step, and native 2-bit quantization lets it fit in 24 GB of memory, so it can run on a single high-end gaming graphics card. It works with llama.cpp, Ollama, LM Studio, Unsloth and vLLM. NVIDIA targets tasks such as web research, writing and code debugging. The smallest member, Bolt E4B, arrives on October 23 as a 4 GB model for voice assistants and embedded interfaces.
Open data
Along with the models, NVIDIA released datasets on Hugging Face covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding, plus reinforcement learning environments in NeMo Gym. That is a useful resource for teams that want to retrain the models on their own data or study how they were trained.
The blog post does not name the license, so it is worth checking the license on the model cards before commercial use. NVIDIA says Super VL reaches accuracy similar to its largest reasoning model with far fewer parameters and is notably faster than other open agent models in its class, and that Bolt ranks first among open models in its class on PinchBench. These rankings are results reported by NVIDIA. Teams would do well to test the models on their own documents and workflows and watch for comparisons from independent researchers.
