Nvidia's new open models are built to run agents on your own PC, not in the cloud.

Three tiers

Nvidia announced Nemotron 3.5 on October 15. Super VL is "an open 120-billion-parameter model" aimed at enterprise agents; its download name suggests 12 billion active parameters. Nvidia says it matches the accuracy of its larger Nemotron 3 Ultra on instruction-following and tool use, with fewer parameters.

Bolt 74B-A9B is the interesting one for anyone running models locally. Nvidia says it uses "a native 2-bit quantization to fit in 24GB of memory" — the 24 GB in its name is the memory budget it was built for. Bolt E4B is a 4 GB model for conversational agents.

Built for your own hardware

The Bolt models target Nvidia's new AI PCs built on its RTX Spark chip, sold through partners including Microsoft and Dell, and its DGX Spark workstations. Bolt 74B already works with the common local-model tools: llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

Nvidia says the models ship "with open weights, open datasets and training recipes." Super VL is also on OpenRouter, ModelScope, and as an Nvidia NIM service.

"Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build." — Kari Briski, VP of Open Models and Software, Nvidia

The claims

Nvidia says Super VL completes tasks faster than models from Anthropic and OpenAI when post-trained on specific agentic tasks. TechCrunch noted that "such claims are standard." Nvidia names Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce as users of Super VL.