The Pulse
NVIDIA Opens Nemotron Models for Always-On AI Agents
NVIDIA released Nemotron 3.5 Super VL and a 24GB version of Nemotron 3.5 Bolt on October 15, 2026, targeting enterprise and local AI agents. A smaller 4GB Bolt model is scheduled for October 23, while Super VL is available through Hugging F
AI.info Team · Written with AI ·
NVIDIA says its new Nemotron 3.5 Super VL completes agentic tasks up to five times faster than the next-best model in its class. Announced October 15, the 120-billion-parameter model is designed for always-on agents that read visual documents, call tools and work through extended business processes.
The release also adds two smaller models for local deployment: Nemotron 3.5 Bolt 74B-A9B, which NVIDIA says fits in 24GB of memory, and Bolt E4B, a 4GB model scheduled to arrive October 23. Super VL and the 24GB Bolt are available now.
One model for text, charts and documents
Super VL combines language and image understanding for tasks such as reading tables, charts, web pages and PDFs. NVIDIA says the model can extract text from a page, interpret a graph and return insights without handing the work between separate text and vision systems.
The company describes Super VL’s accuracy on multimodal agent tasks as comparable to that of its largest reasoning model, with roughly one-fifth as many parameters. NVIDIA also calls it the fastest open agentic model in its class, citing a result of up to five times the throughput of the next-best model. Those are NVIDIA’s reported comparisons; the announcement does not supply detailed test conditions alongside the claims.
Its training totals about 26 trillion tokens across pretraining and continued pretraining. NVIDIA says it then post-trained the model with roughly 3.8 million samples and reinforcement learning in dozens of environments.
Bolt targets 24GB systems and smaller devices
Bolt 74B-A9B uses native 2-bit quantization to fit within 24GB of memory, according to NVIDIA, and is intended to run on RTX systems and DGX Spark alongside other applications. The company says users can run local agent workflows such as web research, writing and code debugging without sending those tasks to the cloud.
The smaller Bolt E4B is aimed at conversational and physical agents, including local voice assistants and embedded interfaces. NVIDIA lists GeForce RTX, Jetson and RTX/DGX Spark systems as target hardware, with availability set for October 23.
Companies are testing specialized deployments
NVIDIA says organizations are adapting Super VL to their own workflows. Scale AI has post-trained it for a vendor-credit finance agent at Skydance and an M&A due-diligence agent at EY; AfterQuery customized it for credit-agreement analysis at Hebbia. Other examples in the announcement include Ramp for financial documents, CrowdStrike for security work and Perplexity for search and agent tasks.
The model’s open weights, datasets and training recipes are intended to let organizations modify and host it themselves. “The model ships with open weights, open datasets and training recipes,” writes Kari Briski in NVIDIA’s announcement. Kari Briski, NVIDIA Blog
Availability spans cloud, workstations and local hardware
NVIDIA lists Super VL as available through Hugging Face, OpenRouter, ModelScope and its NIM service. The company says it runs on DGX Spark, DGX Station, RTX and RTX PRO systems, Jetson devices, data centers and cloud infrastructure. Bolt 74B-A9B is available on Hugging Face, with support listed for tools including llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
The release gives developers a choice between a large multimodal model for longer business workflows and smaller models designed for local use. The practical comparison will depend on how those systems perform on organizations’ own tasks, hardware and data requirements.