On 15 October, NVIDIA announced available Nemotron 3.5 Super VL and Bolt 74B-A9B models. Bolt E4B is scheduled for 23 October, so not every announced variant is available today.

When comparing a local agent with a server-based service, the model and its execution environment should be considered separately. A shared family name does not promise identical accuracy, speed and functionality in every deployment.

Super VL also processes visual documents

The checked NVFP4 model card lists 120 billion total and 12 billion active parameters, text/image/video input and text output. Its stated maximum is one million tokens shared by input, history and output. This is a model capability; a serving recipe can configure a smaller window to accommodate memory and concurrency.

NVIDIA's technical article discusses post-training for tools, organizational procedures and documents. Tables, charts and OCR are among its targets. These capabilities make visual-document tasks relevant, but the company's examples do not constitute SIExpo verification of accuracy on arbitrary material.

In an illustrative task, an agent extracts a value from a chart and connects it to a note elsewhere in a report. A verifiable result needs the correct value, unit and reference to its location. If the chart is read but the note omitted, an answer may sound convincing while remaining unsuitable. This is our evaluation example, not a published NVIDIA test.

Quantization needs its own comparison

The technical report dated 14 October describes combining distillation, long-context training and inference optimization. For NVFP4, its authors report 99.62% median accuracy recovery relative to BF16 across 36 benchmarks. A median over a suite does not mean identical performance on every test. The figure is the developer's report, not an independent guarantee.

Our interpretation keeps model footprint and acceptable error as separate questions. Lower memory use can make a deployment feasible. Whether the resulting output is sufficiently accurate depends on its task and needs evaluation using matched inputs and explicit criteria. Neither assessment substitutes for the other.

A local agent is an entire system

The announcement positions Bolt 74B-A9B for local operation in 24 GB of memory using native 2-bit quantization. That is the intended footprint of this variant, not Super VL's size. A Bolt characteristic should not be transferred to every member of the family.

Even when inference itself is local, an agent may use external search, connected services or remote tools. The weights' location therefore does not describe the complete data flow. A real review would trace what stays on the machine and what is sent during each tool invocation.

For deployment comparisons, we would record the exact checkpoint, numerical precision, serving software and test task. We would then distinguish time to an answer from time to correctly completed work. An agent that generates quickly but performs additional unsuccessful actions can have a different practical cost.

Today's verifiable development is the availability of new artefacts and documentation, with a separately announced smaller future model. SIExpo has not run its own comparative tests. Leadership and acceleration claims remain NVIDIA's and are not presented as universal outcomes for every application.

Editorial verification: 2026-10-15T16:12:02.996741+00:00 (UTC). Cover: AI illustration, not a documentary photograph.