Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models

NVIDIA says it released Nemotron 3.5 Super VL, an open 120-billion-parameter model aimed at specialized always-on agents, claiming fastest task completion and highest accuracy in its class — including up to 5x faster throughput than the next best on Artificial Analysis comparisons NVIDIA cites. Joining it are Nemotron 3.5 Bolt 74B-A9B 24GB for local agentic workloads and Bolt E4B 4GB for conversational and physical agents; Super VL and Bolt 74B are available now, with Bolt E4B due October 23.

NVIDIA says Super VL was built with the Nemotron Coalition, ships open weights, datasets, and recipes, was trained on about 26 trillion tokens then post-trained with roughly 3.8 million samples, and is available on Hugging Face, Open Router, ModelScope, and as NVIDIA NIM. Named customizers and integrators include Scale AI, AfterQuery, Atlassian, CrowdStrike, Perplexity, Salesforce, and others.

This is NVIDIA selling the full agent stack: open models tuned for its GPUs, from DGX down to GeForce. Super VL’s pitch — frontier-ish multimodal agent accuracy at roughly one-fifth the size of NVIDIA’s largest reasoning model — is aimed at enterprises that want to post-train and keep data on-prem.

We’re in favor of open agent models you can customize. The vendor lock-in risk is obvious: “optimized for NVIDIA” is both a performance win and a gravity well. Bolt packing into 24GB via native 2-bit quantization is the part that could matter for developers who refuse the cloud for everyday agent chores.

I’m watching independent Artificial Analysis and PinchBench numbers without the blog framing, and whether Bolt E4B’s October 23 drop actually lands in Jetson and RTX voice-agent builds.

Lightning (August) targets high-throughput; Ultra targets hardest reasoning; Bolt targets the edge. Super VL’s vision focus is OCR, tables, charts, and structured visual documents for agent workflows.