NVIDIA today released Nemotron 3.5 Super VL, a 120-billion-parameter open model for AI agents that work with documents, images and business tasks, alongside a smaller model designed to keep agent workloads local. Nemotron 3.5 Bolt 74B-A9B 24GB is also available now, while Nemotron 3.5 Bolt E4B 4GB is scheduled for October 23.
For users running agents on their own hardware, Nemotron 3.5 Bolt 74B-A9B 24GB uses native 2-bit quantization to fit into 24GB of memory while leaving room for other applications, according to NVIDIA. The company says it can handle web research and code development through local agent harnesses without moving that work to cloud services. It is downloadable through Hugging Face and supports software including LM Studio and Ollama.
Read more: Nvidia Releases GeForce Now Linux App From Beta With Cloud Upgrades
Hardware support for Nemotron 3.5 Bolt 74B-A9B 24GB includes DGX Station, GeForce RTX PCs and DGX Spark, with NVIDIA also targeting RTX Spark laptops scheduled to arrive October 16. NVIDIA is positioning Nemotron 3.5 Bolt E4B 4GB for voice assistants, interactive characters and embedded interfaces on hardware including Jetson and GeForce RTX.
NVIDIA says Nemotron 3.5 Super VL approaches the accuracy of its largest reasoning model while using approximately a fifth as many parameters. Its visual capabilities cover reading charts, extracting text and interpreting tables in webpages and PDFs, allowing one model to handle both visual material and text. Organizations can further train it using their own information and operate it on infrastructure they control.
Nemotron 3.5 Super VL is available through ModelScope community, Hugging Face, NVIDIA NIM and Open Router. NVIDIA is supplying training recipes and datasets alongside the weights, with customization tools that include NVIDIA NeMo. Training involved approximately 26 trillion tokens, followed by post-training using about 3.8 million samples and reinforcement learning.
Companies adapting Nemotron 3.5 Super VL include Ramp for financial documents and R1 for medical claims and revenue cycle management. Scale AI has also post-trained it for EY’s agent handling M&A due diligence and Skydance’s agent handling vendor credit finance.
The releases follow NVIDIA’s September introduction of its Open Agent Safety Platform, which pairs the OpenShell runtime boundary with Sentry, a separate system for monitoring agent behavior. NVIDIA said OpenShell was broadly available, while Sentry uses BlueField-4 DPUs to enforce policies independently of the machine executing an agent.
