Super VL and Bolt
On October 15, NVIDIA released Nemotron 3.5 Super VL, an open 120-billion-parameter multimodal model, alongside Bolt 74B-A9B. The company is publishing Super VL weights, datasets and training recipes. The model is available through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM; Bolt 74B is distributed on Hugging Face and supported by llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
Bolt’s main engineering choice is native two-bit quantization. The 74B-A9B version fits in 24GB of memory, allowing it to run on compatible RTX PCs and compact DGX Spark systems without requiring every workload to send data to the cloud. NVIDIA says the smaller, roughly 4GB Bolt E4B will follow on October 23.
Choosing hardware and a model
The release narrows the gap between server-class multimodal models and local agents: developers can use the larger model for demanding perception tasks and the compact one for private automation on a workstation. That matters to organizations that need controlled local environments, reproducible weights and compatibility with familiar inference tools.
NVIDIA reports fivefold Super VL speedups, up to eight times as many users served and latency reductions of up to 85% in its test configurations. Developers can access the weights, data, training recipes and deployment options.