NVIDIA released Nemotron 3.5 Super VL model weights on October 15. The vision-language model accepts text, images and sampled video frames, and produces text. Developers can use it as the reasoning component of an agent that reads a document or image, calls a tool and continues from the result. The release gives them a model artifact and serving instructions; it does not establish how reliably such an agent performs in a particular workplace.

The big change

  • What changed: NVIDIA has released a post-trained model that combines visual input, text reasoning and tool use in one set of weights. Its documented output is text, so an application still supplies the tools and decides what the model may do.
  • Why it matters: A developer building document or video workflows can evaluate one model across visual evidence and subsequent tool calls, instead of treating visual interpretation as an isolated step. The benefit depends on results with the developer's own material and permissions.
  • What to watch: The released recipes make evaluation possible, but NVIDIA's training guide says some post-training data and components remain unavailable. Independent, task-specific results will say more about the model's usefulness than launch comparisons alone.

What NVIDIA released

The BF16 model card labels the post-trained checkpoint version 1.0 GA and dates its Hugging Face release to October 15. NVIDIA also posted an NVFP4 quantized checkpoint for more compact inference. The model has 120 billion total parameters, with about 12 billion active for each token. NVIDIA's research page appeared on October 14 and describes the broader Super and Lightning models; that earlier research publication is distinct from the October 15 weight release.

The cards list the license as OpenMDW 1.1 and say the checkpoints are ready for commercial and noncommercial use under those terms. The BF16 weights are positioned for customization and full-precision evaluation; the NVFP4 variant is positioned for more efficient serving. NVIDIA's training recipe also lists base weights, data and recipes, but cautions that the supervised fine-tuning and reinforcement-learning data used for Super 3.5 are not yet published. Some teacher and vision-reasoning stages depend on components that are not public. The released materials therefore do not amount to a complete reproduction of NVIDIA's training run.

How an agent can use it

NVIDIA documents text, JPEG or PNG images, and MP4 video of up to two minutes as inputs. Documents can be supplied as page images with accompanying text. The output is a text string, which can contain an answer or a tool call for an application to handle. The model card's API examples show an OpenAI-compatible chat endpoint served locally with vLLM, including examples for images, video and tool calling. This is an interface for developers, not a finished document-processing or incident-response product.

NVIDIA's launch post describes document extraction, chart interpretation and video review as intended applications. It says Ramp is converting receipt images to structured JSON and expects up to 50% more throughput and tenfold lower latency. NVIDIA says Scale AI post-trained the model for vendor-credit processing at Skydance, with a projected 40-fold lower inference cost and fivefold lower latency for that workload. Those are partner-specific claims and expectations in NVIDIA's account, not results measured by BIG CHANGE or evidence that a general deployment will realize the same gains.

Hardware, cost and limits

The BF16 card provides vLLM configurations for four H100 GPUs or two GB200 GPUs. The NVFP4 card lists more compact configurations, including a single DGX Spark for text and images, and a single H200 or GB200 for text, images and video. A maximum context of one million tokens is specified, although the cards warn that individual serving recipes may use a smaller context window to fit the available GPU memory and concurrency. Their video recipe samples frames rather than processing unrestricted footage; its examples start with clips of one to two minutes.

The weights have license terms, but running them still requires suitable GPU capacity and software. The cited model cards do not give a single dollar price for self-hosted inference, and the hosted access link in NVIDIA's launch post did not expose a public price in this review. Cost will depend on the chosen hardware or provider and the actual workload. NVIDIA reports strong benchmark and throughput results, including an up-to-fivefold single-GPU throughput comparison in its launch post. That comparison depends on its stated endpoint, workload and baselines; it is not a measured cost or speed for a reader's application.

Developers evaluating the release should test their own documents, clips, tool permissions and error handling before giving an agent operational authority. NVIDIA's model card itself calls for use-case-specific system testing. BIG CHANGE has reviewed the published artifacts and documentation but has not run the model.

Sources & further reading

  • NVIDIA Technical Blog, October 15: the launch account, proposed use cases, partner claims and vendor throughput comparisons. Partner projections and NVIDIA benchmarks are attributed above.
  • NVIDIA Nemotron research page, October 14: describes the broader model family, research report and linked checkpoints. Its publication date precedes the model cards' release date.
  • BF16 model card and NVFP4 model card, both dated October 15: version, license, inputs and output, serving examples and hardware configurations. The cards are NVIDIA's own documentation, not independent field tests.
  • NVIDIA training recipe: distinguishes the released checkpoints and recipes from post-training data and components still pending public release.