Two months after our August 15 coverage of Nemotron, the family just gained a new, openly released model built for multimodal reasoning and always-on agents.
NVIDIA published Nemotron 3.5 Super VL, an open model that accepts text, images, documents and video and returns text. The company also published technical documentation and made checkpoints available through Hugging Face, OpenRouter, ModelScope and its own NIM service.
Why this matters: until now, the highest-capability multimodal systems lived behind hosted APIs. Making a frontier model and its checkpoints public means teams can inspect, fine-tune and integrate the model directly into tool chains and persistent agents instead of renting access from a closed provider.
How it works, in one image: think of the model as a vast library that never needs to read every book. For each question it walks straight to a few shelves that contain the expertise needed. That architecture lets NVIDIA scale capability without charging full compute for every single request.
What changes now is practical experimentation. Developers and startups can prototype agentic workflows that combine vision, documents and tools on their own infrastructure. NVIDIA also released a smaller Bolt variant intended to run on 24 GB-class systems, which makes local agents more feasible for teams that cannot or will not rely on cloud APIs. Running the full Super VL at scale still requires serious GPU infrastructure, so this is about giving more teams the option to own the stack, not about everyone running it on a laptop.
The open release hands the community a clear test: will independent teams turn these weights into reliable, agentic products that compete with closed systems? Benchmarks and real-world demos should answer that in the months ahead.
