NVIDIA releases Nemotron 3.5 Super VL and a 24 GB Bolt model
Open-weight Nemotron 3.5 Super VL combines text and visual reasoning in one model, while Bolt fits in 24 GB of memory for local agent workloads.
AT A GLANCE
Key points
- Super VL has 120 billion total parameters and activates 12 billion per token.
- Bolt uses native 2-bit quantization to fit in 24 GB of memory.
- Both releases target multimodal, tool-using agent workflows.
What is changing?
NVIDIA released the open and customizable Nemotron 3.5 Super VL model on October 15. Its 120-billion-parameter mixture-of-experts architecture activates 12 billion parameters per token. The model combines reasoning, instruction following, and tool use with text, image, document, and video understanding in a single system.
The locally focused Nemotron 3.5 Bolt 74B-A9B 24GB model is also available. According to NVIDIA’s model card, Bolt uses native 2-bit quantization to fit 74 billion total parameters and nine billion active parameters into 24 GB of memory. It is designed to run tool calling, web browsing, and terminal tasks on-device.
Base, post-trained, and NVFP4 versions of Super VL were released alongside training recipes and datasets. Bolt is available through Hugging Face with support for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. The provider’s performance claims should be independently validated across different hardware and workloads.
Explore the original sources
Source date: 15 October 2026.
Published by AI Radar: 16 Oct 2026, 14:16 (Istanbul, UTC+3).
FOLLOW THE TOPIC
Topics and tags
Share your discovery.
Send this story to a friend.
Reader comments
What do you think about this development? Share your experience or question. Comments are published after editorial approval.
Your name/nickname and comment become public if approved. Read: Privacy notice.
Loading comments…