NVIDIA: Nemotron 3.5 Super VL

NVIDIA Nemotron 3.5 Super VL is an open multimodal mixture-of-experts model from NVIDIA. It has 12B active parameters out of 120B total, so only a fraction of the network is engaged per token while the full model retains broad capacity. It is listed on api14 under the text category, and the provider release date is 2026-10-15.

The model accepts text, images, and video as input, and it returns text. That combination covers documents, GUI screenshots, and video alongside ordinary prompts, so a single request can mix written instructions with visual material. The context window is 1,048,576 tokens, and the maximum output is 235,929 tokens. Long inputs are therefore practical: multi-document bundles, extended screenshot sequences, or long video-derived context can fit without aggressive trimming, and the large output ceiling suits tasks that need a substantial written result rather than a short answer.

Those capabilities point to a few concrete uses. Screenshot understanding is a natural fit for GUI agents, interface testing, and support workflows where the input is a picture of a screen rather than structured data. Document handling suits extraction, summarisation, and question answering over scanned or layout-heavy files. Video input supports description, timeline-level questions, and content review. Because the model also handles plain text, it can serve as a single endpoint for pipelines that would otherwise split vision and language work across two systems.

You call it through api14's OpenAI-compatible API using the model id "nvidia/nemotron-3.5-super-vl". The supported request parameters are frequency_penalty, include_reasoning, logit_bias, logprobs, max_tokens, min_p, presence_penalty, reasoning, repetition_penalty, response_format, seed, stop, structured_outputs, temperature, tool_choice, tools, top_k, top_logprobs, and top_p. Tool calling and structured outputs are both available, which matters if the model sits inside an agent loop or needs to emit machine-readable results. Reasoning can be requested and surfaced through include_reasoning, and sampling can be tuned with the usual temperature, top_p, top_k, and penalty settings.

The main constraint to keep in mind is that output is text only, so the model cannot generate images or video. If your task is purely text and does not benefit from visual input, a smaller text-only model may be simpler and faster to operate. If you need image or video generation rather than understanding, this model is not the right fit. For mixed text, image, and video understanding with a very large context, it is a reasonable default.