NVIDIA releases Nemotron 3.5 Super VL and Bolt 24GB; smaller Bolt follows October 23
NVIDIA released Nemotron 3.5 Super VL and Nemotron 3.5 Bolt 74B-A9B 24GB on October 15. The company says a smaller Bolt E4B 4GB model is due October 23. The announcement therefore does not make all three models immediately available.
Super VL is described as an open 120-billion-parameter vision-language model for agentic work. NVIDIA highlights instruction following, tool use, multistep workflows and understanding text, images, charts and documents. These capabilities are the vendor's description, not a guarantee of correct output on every task.
NVIDIA says Super VL ships with open weights, datasets and training recipes and can be adapted through its NeMo tools. Availability is listed through Hugging Face, OpenRouter, ModelScope and NVIDIA NIM. Buyers should check the actual distribution's license, deployment requirements and service terms rather than treat open weights as an unlimited hosted service.
Bolt 74B-A9B 24GB targets local agentic workloads on NVIDIA systems and uses native 2-bit quantization to fit within a 24GB model-memory footprint. The release names support for llama.cpp, Ollama, LM Studio, Unsloth and vLLM. A model-memory figure does not establish the total memory required by every application or context size.
The coming Bolt E4B 4GB model is aimed at conversational and physical agents on GeForce RTX, Jetson and RTX/DGX Spark. NVIDIA also reports speed and benchmark advantages for the family. Those comparisons depend on the tested model class and serving configuration and are not independent measurements of a buyer's workload.
What to check before deployment: the model actually available today, supported runtime, hardware, memory overhead and license. Local execution can reduce the need to send a model workload to a remote provider, but an agent's connected tools may still transmit information. Open weights alone do not establish the privacy or security of an entire application.