Summary
NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter model for multimodal agent tasks, alongside two Bolt models designed for local and edge systems. The company says the models cover deployments from data centres to desktop hardware.
NVIDIA released Nemotron 3.5 Super VL on Oct. 15, 2026, describing it as an open 120-billion-parameter model for multimodal agent tasks. It is joined by two Nemotron 3.5 Bolt models aimed at running agentic workloads on local NVIDIA hardware: Bolt 74B-A9B 24GB, available now, and Bolt E4B 4GB, scheduled for Oct. 23.
The models are intended to span different deployment settings, from enterprise servers to desktop and edge devices. NVIDIA says Super VL combines document and visual understanding with instruction-following and tool use, while the Bolt models prioritise running agents with lower memory requirements.
A model for documents and multi-step tasks
Super VL is designed to handle tasks such as reading text in images, extracting information from tables and charts, interpreting webpages and PDFs, and using tools to carry out multi-step workflows. In an agent system, a model can use this understanding to act across stages of a task—for example, extracting information from a document and then using it in a subsequent operation.
NVIDIA says Super VL delivers accuracy on multimodal agent tasks comparable with its largest reasoning model at roughly one-fifth the parameter count. The company also reports up to five times faster throughput than other models in its class, citing the Artificial Analysis Intelligence Index leaderboard. These are NVIDIA-reported comparisons; the announcement does not set out the test configuration.
The model was trained on about 26 trillion tokens across pretraining and continued pretraining, then post-trained with roughly 3.8 million samples and reinforcement learning across dozens of environments, according to NVIDIA. The company says it is releasing open weights, datasets and training recipes. It also identifies NVIDIA NeMo, Python-based pipelines and open datasets as resources for organisations that want to adapt the model to their own workflows.
Bolt targets local and edge hardware
Bolt 74B-A9B 24GB is designed for everyday agent tasks on NVIDIA RTX and DGX Spark systems. NVIDIA says its native 2-bit quantization allows it to fit in 24GB of memory and run alongside other applications. Quantization represents model weights with fewer bits, reducing memory use; the company positions this version as a balance between capability and speed for local workloads.
NVIDIA says the model can be used with local agent frameworks for tasks including web research, writing and code debugging, without sending those workloads to the cloud. It is available on Hugging Face, with support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM.
Bolt E4B 4GB is aimed at conversational and physical agents, including local voice assistants, interactive characters and embedded interfaces. NVIDIA lists GeForce RTX, Jetson and RTX/DGX Spark among the systems it is designed for.
Super VL is available through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM, and is intended to run across NVIDIA systems including DGX, RTX and Jetson hardware as well as data-centre and cloud environments. Together, the releases give developers model options for different memory and deployment constraints, while the open weights and training resources allow organisations to adapt the models rather than use them only as fixed hosted services.