NVIDIA launches Nemotron 3.5 Super VL for multimodal agents

A multimodal model built for specialised agents

NVIDIA has introduced Nemotron 3.5 Super VL as an open, customisable foundation for agents that reason over text, images, documents and video. The mixture-of-experts model contains 120 billion parameters but activates about 12 billion for each token. That structure aims to combine broad capacity with a lower inference burden than a dense model of comparable total size.

The release targets more than visual question answering. NVIDIA positions it for tool use, document workflows, video understanding and multi-step agent tasks in which visual evidence must inform an action. Enterprises can post-train the model for a domain rather than sending every task to a closed general-purpose service, but that flexibility also transfers evaluation and operational responsibility to the adopter.

Throughput claims need workload context

NVIDIA says Super VL can deliver up to five times the throughput of open peers on a single GPU. Such figures are useful for screening, but they depend on hardware, precision, input mix, context length and batching. A document-heavy agent that processes high-resolution pages may behave very differently from a short text-and-image benchmark.

Teams should measure completed business tasks per dollar and per unit of latency, not just tokens per second. They should also report quality at the precision and serving configuration used in production. A quantised path that increases throughput but misses small visual details may be a poor trade for insurance forms, manufacturing inspection or compliance review.

Post-training is the main enterprise proposition

NVIDIA highlights work with Scale AI, Ramp and AI21 Labs to demonstrate specialised post-training. The published examples include claims of up to 40 times lower cost, ten times lower latency and a 40 per cent accuracy improvement in selected settings. These are customer-specific results rather than guarantees, but they illustrate the intended pattern: start with the open base, then adapt it to a bounded workflow and deployment environment.

A credible adaptation program needs representative data, held-out evaluation and protection against leakage. Visual training sets can contain customer records, faces, documents and location details that require careful handling. Teams should record the base checkpoint, training recipe and data lineage so a performance gain can be reproduced and a harmful behaviour traced to its source.

Multimodal agents broaden the attack surface

Images and documents are not passive inputs. They can contain instructions intended to redirect an agent, conceal text from a human reviewer or exploit weaknesses in downstream parsers. Video adds temporal ambiguity and much larger evidence volumes. An agent that can also call tools may convert a misread visual cue into an external action.

Operators should isolate retrieved content from trusted instructions, scan files, restrict tool permissions and require approval for consequential steps. Evaluation needs adversarial documents and images as well as ordinary accuracy cases. The system should cite the region, page or frame supporting a decision so reviewers can tell whether the model relied on genuine evidence or an irrelevant visual pattern.

Where the release could matter

An open multimodal model can help organisations keep sensitive inputs inside a chosen environment, tune terminology and integrate with established software. The 12-billion-active design may make that control practical on less infrastructure than the total parameter count suggests. It also gives researchers a base they can inspect and compare rather than an opaque endpoint.

The decisive evidence will come from independent tests across messy documents, long video, tool execution and safety boundaries. NVIDIA has provided a strong technical and commercial claim, not a universal result. Adopters should begin with a narrow task, measure both accuracy and action safety, and expand only when the adapted model consistently outperforms simpler systems on the workflow that actually matters.

Choosing between adaptation and retrieval

Not every specialised requirement calls for post-training. Retrieval can supply current product manuals or policies without changing model weights, while fine-tuning may be better for stable terminology, output structure or repeated visual patterns. A team should compare the approaches against the same held-out cases and consider how quickly knowledge changes. Embedding volatile facts in weights can create an expensive update cycle and make provenance harder to explain.

A layered design may use retrieval for evidence, a small adapter for domain behaviour and deterministic tools for actions. That separation makes it easier to update one component and to identify where an error began. It also prevents a strong benchmark result from encouraging an oversized model to perform work that rules-based software can handle more reliably.

Operational takeaway

Procurement decisions should include the full serving stack. A permissive or open model can still depend on proprietary accelerators, optimised libraries and specialist engineering. Teams should calculate the cost of post-training, monitoring, upgrades and fallback capacity alongside inference. They should also test portability before committing to a production architecture. If the model meets quality goals only with one particular stack, that may be acceptable, but the dependency should be visible. The value of customisation is greatest when an organisation can reproduce its system and understand which component created the improvement.