NVIDIA has added three new open models to its Nemotron lineup, targeting the full spectrum of AI agent deployments from data centers to desktops. The headliner is Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model engineered for specialized, always-on enterprise agents.
Super VL uses a mixture-of-experts architecture that activates only 12 billion parameters per token. NVIDIA claims accuracy comparable to its largest reasoning models at roughly one-fifth the parameter count, with up to five times the throughput of other open models in its class. The model handles instruction following, tool use, multistep business workflows, and visual understanding of documents, charts, and PDFs. It was trained on about 26 trillion tokens with reinforcement learning across 100 open environments, and NVIDIA is releasing the weights plus open datasets on Hugging Face so enterprises can fine-tune it on proprietary data behind their own firewalls.
Two smaller siblings cover local deployments: Bolt 74B-A9B 24GB, which fits in 24GB of GPU memory for local agent workloads on RTX hardware and is available now, and Bolt E4B 4GB, a compact conversational model arriving October 23. Early adopters include Scale AI, which customized Super VL for financial vendor processing at Skydance with a reported 40x inference-cost reduction, and Ramp, which uses it for receipt and invoice parsing.
The release was built with the Nemotron Coalition, whose members include Mistral AI, Thinking Machines Lab, and Black Forest Labs, with Perplexity among the platform partners integrating the model. Developers comparing agentic coding setups can also check Claude Code and Cursor, or read our AI tool buyer’s guide for choosing the right stack.
Sources: SiliconANGLE, The Decoder
