NVIDIA released Nemotron 3.5 Super VL on 15 October, a model that reads text, images, documents and video and is built to carry out long business tasks on its own. The company’s central claim is speed. A chart in its developer blog post says higher throughput and “token efficiency” let the model finish agent tasks four times faster.

Token efficiency is the part that touches a budget. Models that reason write out their thinking as they go, and most hosted services charge for every token produced. A model that reaches the same answer in fewer steps costs less to run and returns sooner. According to the post, completion time depends on how many tokens get used, how often the agent calls tools, and how many moves it takes to land on the right result, but it publishes no token counts. The 4x figure also comes without a named comparison, so the reader cannot tell what it beats.

The same goes for a second claim: up to 5x higher throughput on a single chip “compared to other open models in its size category.” No rival is named. NVIDIA also reports more than 600 tokens per second on Artificial Analysis’s agentic tasks, measured on the hosting company Baseten’s service. Artificial Analysis is an outside testing firm, but that line is a speed measurement run on its tasks. The post gives no score from the firm’s Intelligence Index and no ranking. A reader who has seen a 26.5 attached to this model should know it is not in NVIDIA’s account.

On accuracy, NVIDIA compares the model with its own larger Nemotron 3 Ultra. Super VL scores 63.44% on Hermes Bench against 56.33%, and 87.05% on LangChain’s evaluations against 84.3%. Both tests come from agent-software makers, but the post does not say who ran these particular numbers. The one outside check is from EXO Labs, which it says placed the model on the “Pareto frontier” (the best available trade-off between accuracy and speed) when run locally on NVIDIA’s DGX Spark desktop machine.

The model has 120 billion parameters in a mixture-of-experts design, which means only 12 billion are active for any given token. That is how a large model stays cheap per word. NVIDIA says it was distilled from Ultra, meaning trained to imitate the larger system. The post does not mention a smaller “Bolt” model. It names Nemotron 3.5 Lightning as the small text model for high-volume work.

The compressed version is the least documented piece. NVFP4 stores weights in four bits instead of the usual sixteen, and NVIDIA says a training step lets the shrunken model “match the outputs of the original” so accuracy holds. The post gives no accuracy loss and no file size. On NVIDIA’s newest Blackwell chips the 4-bit format applies to the calculations as well. On older hardware, the post says only weights are 4-bit.

Customer figures are mostly projections. Scale AI says a vendor-credit workload for Skydance “will decrease” in cost 40x. Ramp “expects” up to 50% more throughput. AI21 and Alice report 41% better accuracy on threat analysis after retraining the model.

On access, NVIDIA calls the model open and says organizations can adapt and deploy its open weights. The release covers base, post-trained and NVFP4 checkpoints, training data and recipes, available through Hugging Face, ModelScope and OpenRouter. The post does not state a licence or its terms, so the conditions for commercial use need checking before anyone builds on it.

NVIDIA sells the chips this model is tuned for. The post leans on Blackwell, B200 and DGX Spark results, which is a fact about who is publishing and a reason to rerun any speed claim on the hardware you will use.

Anyone paying per token should run their own workload through the model and count the output, because that figure is the one NVIDIA did not publish.

Based on NVIDIA developer blog post of 15 October 2026