NVIDIA puts vision inside its open Nemotron line: 120B parameters, 12B active per token, GDPval Elo up from 493 to 1190
Synopsis
NVIDIA released Nemotron 3.5 Super VL, an open multimodal model now available for inference and training on Baseten, so agents can read documents, interpret application screens and analyze video directly. It builds on the Nemotron 3 Super architecture as a 120B-parameter Mixture-of-Experts model that activates 12B parameters per token, adding a vision encoder for image input. In NVIDIA's reported evaluations it scores an Elo of 1190 on GDPval versus 493 for Nemotron 3 Super, and also scores higher on SciCode and GPQA-diamond.
Interpretation
The open Nemotron family gains a multimodal member that handles visual input directly, letting an agent inspect a document layout, interpret a screenshot or analyze a video clip and then reason about the next step. Previously the family's text model, Nemotron 3 Super, could only take visual information after it had been converted to text. NVIDIA's reported evaluations show text reasoning did not regress when vision was added: Nemotron 3.5 Super VL earns an Elo of 1190 on GDPval against 493 for Nemotron 3 Super, and also scores higher on SciCode and GPQA-diamond.
A vision encoder is added to the existing architecture so images and text enter one model, which decides the next step and generates tool calls the application can execute. A hybrid Mamba-Transformer architecture supports long-context processing, multi-token prediction (generating several subsequent tokens at once to speed output) accelerates generation, and NVFP4 KV cache quantization is designed to cut memory use for long-context inference. On Baseten, developers can evaluate the model with NVFP4 weights on NVIDIA Blackwell GPUs and adapt it to their own data through Baseten Loops for supervised fine-tuning and reinforcement learning, then deploy the trained checkpoint on the same platform.
Perspective
The release targets developers whose applications handle documents, images, interfaces and video, and covers two paths: inference evaluation and training customization on Baseten. NVIDIA's comparison is confined to three text benchmarks, used to show that text reasoning holds up when vision is added. The release suggests developers first measure, through the Model API, how well the model understands their own images, documents or screenshots, how reliably it produces the needed outputs, and whether latency fits their workflow, then connect application tools and post-train on their own data with Loops.
The reported comparison lists only three text benchmarks, GDPval, SciCode and GPQA-diamond, and gives no quantitative results on vision tasks. The GDPval figures of 1190 and 493 are Elo ratings whose comparability depends on the same evaluation setup. How the model performs on specific document, interface or video workflows still needs to be measured with a developer's own inputs.