NVIDIA Nemotron 3.5 Super VL and Bolt: Open Models for AI Agents
On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter multimodal MoE model with 12B active parameters. It also released Nemotron 3.5 Bolt 74B-A9B, which fits in 24GB of memory so agents can run locally.
Super VL is a hybrid Mamba-Transformer MoE with a context window of up to 1M tokens. It’s built for always-on, specialized enterprise agents that reason over text, documents, images and video. NVIDIA is releasing base, post-trained and NVFP4 checkpoints along with training recipes, RL environments and open datasets for SFT and RL. Teams can post-train it on their own data with NeMo.
Bolt 74B-A9B is a compressed version of Super VL. It uses native 2-bit quantization to fit 24GB, and it supports llama.cpp, Ollama, LM Studio, Unsloth and vLLM. It targets RTX Spark laptops, GeForce RTX PCs and DGX Spark. A smaller Bolt E4B 4GB model is due October 23.
Super VL is available on Hugging Face, OpenRouter, ModelScope and NVIDIA NIM.
FAQ
Can Nemotron 3.5 run on a local PC?
Yes. Bolt 74B-A9B fits in 24GB and runs on GeForce RTX PCs and DGX Spark.
Is Nemotron 3.5 Super VL open for fine-tuning?
Yes. NVIDIA released its checkpoints, datasets, recipes and RL environments.