NVIDIA has released Nemotron 3.5 Super VL, a 120-billion-parameter mixture-of-experts model with only 12 billion parameters active at any one time. It reads text, images, documents and video, and is built specifically for agentic work: reasoning through a task and calling tools rather than just answering a single question.
NVIDIA's own comparison claims Super VL matches its largest reasoning model on multimodal agent benchmarks while using around a fifth of the parameters, and that it delivers up to five times the agentic throughput on a single GPU compared with open models of similar size. Independent tracking from Artificial Analysis places the model at 26.5 on its Intelligence Index, giving outside parties a figure to check NVIDIA's claims against over time.
The model was developed with the Nemotron Coalition and launches alongside a smaller sibling, Bolt 74B-A9B, also available now; a further edge-focused variant, Bolt E4B, is due on 23 October. NVIDIA is releasing base, post-trained and NVFP4-quantised checkpoints for Super VL, together with the training recipes, data and reinforcement-learning environments used to build it, a more complete release than a weights-only drop.
For teams building multimodal agents, that combination of a smaller active-parameter count and openly published training material matters more than raw benchmark position: it makes the model cheaper to serve and easier to adapt than a closed, single-checkpoint release.
Source: MarkTechPost
