Nemotron 3.5 Super VL

NVIDIA has launched Nemotron 3.5 Super VL, its open multimodal model for specialized agents. It is the company’s 33rd launch on Product Hunt, and this one goes after a very specific job: letting agents handle more than plain text without juggling separate systems.

The model is built as a 120B mixture-of-experts system with 12B active parameters. That matters because NVIDIA is pitching it as one model that can take text, images, documents, and video, then do reasoning and tool use on top. In other words, the whole stack is meant to feel less stitched together.

NVIDIA says the model can deliver up to 5x higher agentic throughput on a single GPU than open models of the same size. That is the kind of line that will get attention fast, especially from teams trying to squeeze more work out of one machine instead of scaling everything out.

The release is also unusually open for a hardware giant. Checkpoints, training recipes, and datasets are available on Hugging Face, along with a NIM container. So this is not just a demo with a glossy label; NVIDIA is handing over the pieces people would need to inspect, adapt, and build from.

The model is free, which is the kind of word that never hurts a launch. But the real story is simpler: NVIDIA is trying to make its own open model the place where multimodal agents start, not the place where they get bolted together later.