Nemotron 3.5 Super VL needs workload evidence beyond its throughput headline

NVIDIA introduced Nemotron 3.5 Super VL on 15 October 2026 as a 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token. NVIDIA claims up to five times the single-GPU throughput of comparable open models on selected agent tasks. The numbers make the release worth testing, but they do not establish the cost or reliability of a deployed multimodal agent.

The model combines text, documents, images, video understanding, reasoning and tool use. That can remove separate vision-language hand-offs in workflows such as document extraction and threat analysis. It can also bring more kinds of untrusted evidence into a single decision path. Fewer components are not automatically fewer security boundaries.

Model size is not the serving bill

NVIDIA reports that the model's mixture-of-experts design activates only part of its parameter capacity for each token. It also describes NVFP4 quantisation, speculative decoding and context reuse across agent turns. Each technique targets a real inference cost, but a headline token throughput result excludes the full expense of video processing, retrieval, tool calls, retries and human corrections.

NVIDIA reports 63.44% on Hermes Bench and 87.05% on LangChain evaluations for Super VL, compared with 56.33% and 84.3% for its Nemotron 3 Ultra in those evaluations. Operators should treat these as task-specific vendor-reported figures. Reproduce acceptance tests using their own prompts, representative data, serving hardware and tool configuration rather than assuming the ranking transfers.

An open multimodal model earns production capacity through verified task outcomes, not the fastest token rate on a vendor chart.

Customisation changes the assurance boundary

NVIDIA points to domain post-training with NeMo, including supervised fine-tuning, low-rank adaptation and reinforcement learning. Its reported customer examples include a 40% accuracy improvement in one threat-intelligence setting and major efficiency gains in document workflows. Those results depend on the specific training and evaluation conditions; they are not universal properties of the checkpoint.

For security-oriented use, maintain a frozen evaluation set covering correct extraction, false relationships, deceptive visual material, sensitive-document handling and unauthorised tool actions. Version the model weights, fine-tuning data, prompt policy and inference configuration together. Re-run the same tests after quantisation or domain adaptation. A change in latency is not acceptable if it silently changes the decision boundary.

Local deployment needs independent controls

NVIDIA lists support for local, data-centre and cloud deployments, including DGX Spark, NIM and Dynamo configurations. Open weights can give operators greater control over placement and data flows. However, hardware dependencies, serving licences, audit telemetry and access policies still shape portability in practice.

Before adopting Super VL, compare it against the incumbent model on cost per accepted workflow, evidence traceability and tool-call errors. Enforce authorisation outside model reasoning and isolate confidential multimodal inputs. Record memory pressure, cache retention and recovery behaviour alongside model accuracy. The release broadens deployment choices; only workload-level evidence can justify replacing an established production system.

Produced through Devence Lab's automated research and publishing workflow, without prepublication human review. Devence Lab retains editorial responsibility. Sources are linked; report factual errors to contact@devencelab.com.

Written by the Devence Lab research team.