NVIDIA’s 120-billion-parameter agent activates 12 billion at a time—and Bolt fits into 24 GB

NVIDIA has split its newest open agent models across two very different deployment problems. Nemotron 3.5 Super VL is a 120-billion-parameter vision-language model intended for high-volume agents that read documents, use tools and work across long contexts. Nemotron 3.5 Bolt 74B-A9B is a compressed descendant intended to run locally inside a 24 GB memory budget.[1][2][3]

The numbers sound contradictory only if every parameter is assumed to participate in every answer. Super VL is a mixture-of-experts model: it stores 120 billion parameters but activates about 12 billion for each token. The model card also gives it a context limit of up to one million tokens and support for text, images and video.[2]

Compression is the product, not an afterthought

Bolt was not trained as an unrelated small model. NVIDIA says it compressed the 120B-A12B parent to 74 billion total and 9 billion active parameters, then applied further post-training, roughly two-bit expert-weight quantization and speculative decoding. The design question was therefore: how much of a large model’s behaviour can survive inside a fixed local memory envelope?[3]

That trade matters because a local agent competes with the rest of the machine. A model that technically fits only after consuming all available memory leaves little room for the browser, code editor, retrieval index or other applications that make the agent useful. The 24 GB target is an operating constraint, not merely a download-size headline.

The benchmark claim needs its boundary

NVIDIA describes Super VL as up to five times faster than the next model in its class and Bolt as highly competitive with models that occupy similar memory. Those comparisons are useful launch evidence, but they are still produced by the vendor, depend on selected classes and hardware, and do not establish the best model for every agent workload.[1][3]

Open weights change what can be controlled

NVIDIA is releasing weights, datasets and training recipes for Super VL, and checkpoints for Bolt. That lets organisations post-train, inspect deployment behaviour and keep sensitive workloads on infrastructure they control. Openness does not remove security or evaluation work; it moves more of that responsibility to the deployer.[1][3]

The useful headline is therefore not simply that NVIDIA released another 120-billion-parameter model. It is that one sparse parent is being offered in two forms: a large multimodal agent model and a deliberately compressed local derivative. The engineering story is the route between them.

Report an error or suggest a correction