NVIDIA Expands Nemotron AI Models for Data Centres and Desktop Devices

NVIDIA announced three additions to its Nemotron family of open AI models. The models target enterprise workflows, local AI agents and conversational applications across data centres, desktop systems and edge devices. 

The launch includes Nemotron 3.5 Super VL, a 120-billion-parameter model, and two smaller models under the Nemotron 3.5 Bolt series.

NVIDIA said Super VL delivers up to 5x the throughput of the next-best model in its class. Nemotron 3.5 Bolt 74B-A9B 24GB and Nemotron 3.5 Bolt E4B 4GB target local AI workloads. 

Super VL and Bolt 74B-A9B are available now, while Bolt E4B is scheduled for release on October 23.

Super VL supports tasks such as instruction following, tool use, multi-step business workflows and visual understanding. NVIDIA said the model achieves accuracy comparable to its largest reasoning model with roughly one-fifth of the parameters.

This lets organisations customise Super VL with their own data, deploy it on their infrastructure, and retain control over their applications. The model can process webpages, PDFs, tables and charts, extracting information for AI agents to use.

NVIDIA has developed the model with the Nemotron Coalition. It was trained on approximately 26 trillion tokens across pre-training and continued pre-training. The company then post-trained it using roughly 3.8 million samples and reinforcement learning across dozens of environments.

Nemotron 3.5 Bolt 74B-A9B 24GB is designed for everyday AI agent tasks on NVIDIA RTX systems and DGX Spark. It uses 2-bit quantisation to fit within 24GB of memory and run alongside other applications.

Users can use the model for tasks such as web research, writing and code debugging without sending workloads to the cloud. NVIDIA said the model ranks first among open models in its class on PinchBench and second on BrowseComp.

The smaller Nemotron 3.5 Bolt E4B 4GB targets conversational and physical AI agents. NVIDIA designed it for GeForce RTX, Jetson and RTX/DGX Spark systems. It supports local voice assistants, interactive characters and embedded interfaces.

NVIDIA said Scale AI has post-trained the model for Skydance’s vendor credit finance agent and EY’s mergers and acquisitions due diligence agent.

Other applications include credit agreement analysis for Hebbia, cybersecurity through AI21 and Alice, and coding through Mercor. Ramp is customising the model for financial document intelligence, while Lila Sciences is using it for scientific discovery and search.

Software companies are also integrating Super VL into their products. Atlassian is exploring agentic search across Jira and Confluence, while CrowdStrike is customising it for enterprise security operations. 

Perplexity is integrating the model for search and AI agent tasks.

NVIDIA said organisations can deploy Super VL according to their security and data residency requirements. The model includes open weights, datasets and training recipes. Developers can customise it through NVIDIA NeMo and Python-based pipelines.

Super VL is available through Hugging Face, OpenRouter, the ModelScope community and NVIDIA NIM. The Bolt 74B-A9B model is also available on Hugging Face, with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.