NVIDIA ships Nemotron 3.5 Super VL open-weight model

NVIDIA positions Nemotron 3.5 Super VL as a model built for AI agents that combine multiple pieces of information. Its Mixture of Experts (a design that activates only part of the model for each task) keeps the active parameter count at 12 billion out of 120 billion, which underpins the throughput claims NVIDIA makes for running agent workloads on a single GPU. NVIDIA reports that on four NVIDIA B200 GPUs the model completed 10,000 software development tasks in about 10 hours at roughly 51% accuracy, up to four times faster than comparable models.

Beyond raw speed, the release leans on fine-tuning. Using NVIDIA NeMo, customers can adapt the model to their own documents and processes, and the examples NVIDIA cites come from that setup: Scale AI's work for Skydance Corporation on unpaid invoice and refund documents, Ramp's receipt extraction, and AI21 and Alice's cybersecurity threat analysis. Each is framed as a tailored deployment rather than an off-the-shelf one.

Supporting techniques aim at operating cost. LiLiCorr, a look-ahead method for generating text, is said to make generation up to 2.3 times faster on NVIDIA's DGX Spark small AI computer, and the model can reuse earlier computations when an agent rereads the same instructions or conversation history. For many simultaneous users, NVIDIA Dynamo is offered to route processing; a test with 16 NVIDIA B200 GPUs is reported to support up to eight times as many users at the same per-person generation speed, with time to first answer shortened by 85%.