NVIDIA releases an open 120B model for always-on agents

NVIDIA published Nemotron 3.5 Super VL on October 15, 2026 and described it as an open model with 120 billion parameters, aimed at specialized agents that stay running. Two smaller edge models were announced beside it. Nemotron 3.5 Bolt 74B-A9B is a 24 gigabyte build meant to balance capability and speed for local agent work, and NVIDIA says it is available now along with Super VL. A smaller Bolt E4B, about 4 gigabytes, is scheduled for October 23.

The company says the three models are meant to cover agents from a data-center machine down to a desk or an edge box. Super VL is listed for DGX Spark, DGX Station, RTX PRO and RTX PCs, Jetson, and cloud or data-center hardware. The weights are posted for Hugging Face, OpenRouter, and ModelScope, and the model is also offered as an NVIDIA NIM. NVIDIA says Super VL was built with the Nemotron Coalition and is intended for instruction following, tool use, multi-step business workflows, and visual understanding. Those are the company's claims about the release.

Why it matters

Teams that cannot send every agent call to a closed API now have another open weight to try on their own hardware.

Source: NVIDIA