NVIDIA opens Nemotron models for visual documents and local agents
NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B, offering visual and text inputs for enterprise workflows and local agents. Super VL handles documents, tables and charts; Bolt targets tasks running on local hardware. Weights and customization resources are available now. The smaller Bolt E4B is scheduled for October 23, with independent evaluation already published for the released models.
Artificial Intelligence··Night
Super VL reads visual and written material together
NVIDIA released Nemotron 3.5 Super VL and Bolt 74B-A9B on October 15. Both models accept images and text and return text. Super VL is aimed at enterprise agents working through documents, charts, tables and web pages; Bolt is designed for multistep agent workloads running on local hardware.[1], [2]
Super VL has 120 billion total parameters, with only a subset active at each step in its mixture architecture. Its stated tasks include optical character recognition, document analysis and tool use. Visual and written information can stay within the same model as an agent moves between a PDF document, its tables and related web material.[1]
Bolt offers a local execution path
NVIDIA provides a native two-bit quantized Bolt version for local NVIDIA systems. Quantization reduces the numerical precision used to represent model weights. The announced support includes llama.cpp, Ollama, LM Studio, Unsloth and vLLM, software used to run or adapt models. Super VL is available through the Hugging Face and ModelScope model repositories and NVIDIA NIM’s deployment service.[1]
Weights arrive now; the smaller E4B follows later
The release includes weights, datasets and training recipes for customization. NVIDIA’s NeMo tools, Python-based pipelines and NeMo Gym environments support that work. The smaller Bolt E4B, announced for physical and conversational agents, is scheduled for October 23. The available Super VL and Bolt release therefore precedes the smaller model’s planned distribution.[1]
Artificial Analysis, an independent model-evaluation organization, published its own assessment of the released models. Its results describe performance on a defined evaluation index. NVIDIA’s specialized speed and task comparisons retain their supplier attribution; the availability of weights and training materials supplies a separate, concrete route for organizations to customize the models.[1]
