Nvidia released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt, a pair of open-weight models aimed at always-on AI agents from the data center down to a laptop. Super VL has 120 billion total parameters (12 billion active), reads text, images, and video, and supports up to a 1 million-token context. Nvidia says it is built for long-running agent workflows and can be post-trained on a company’s own data.
Bolt 74B-A9B uses native 2-bit weights to fit in 24GB of memory for local agent work on GeForce RTX PCs, DGX Spark, and the RTX Spark laptops Nvidia says arrive October 16. A smaller Bolt E4B 4GB model for conversational and edge agents is due October 23. Weights for Super VL and the 24GB Bolt are on Hugging Face now, with Super VL also available as an NVIDIA NIM and on OpenRouter.
The release lands the same day Meta shipped Muse Spark Open for Mac Studio desktops. Nvidia is pitching Nemotron as the open stack for shops that want agent software on their own silicon instead of a closed API.
— via NVIDIA