NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model

NVIDIA has released Nemotron 3.5 Super VL, an open multimodal model aimed at enterprise agents that need to stay on all day and handle more than plain text. The model takes in text, images, documents and video, then reasons, calls tools and writes code from the same checkpoint.

The headline number is 120B total parameters, with 12B active per token through a mixture-of-experts design. NVIDIA says the model supports up to 1 million tokens of context, and that the NVFP4 checkpoint can run on a single DGX Spark, a single H200, two H100s or a GB200. A BF16 checkpoint is also available.

The company is also pushing the packaging as much as the model itself. The weights are under the OpenMDW-1.1 license, and NVIDIA says it has published open datasets, training recipes and reinforcement-learning environments in NeMo Gym. The model card lists text, image and video inputs with text output, while video input is limited to MP4 clips of up to 2 minutes. Reasoning can be switched on or off through the chat template.

Under the hood, NVIDIA says the backbone mixes Mamba-2, mixture-of-experts and attention layers, with 512 routed experts in each MoE layer. It also uses a C-RADIO vision encoder and connector for images, document pages and video frames. Speed comes from several tricks, including multi-token prediction, LiLiCorr on DGX Spark, NVFP4 quantization-aware distillation and prefix caching for agent turns.

On NVIDIA-reported benchmarks from the NVFP4 checkpoint, the model scores 89.90 on GPQA-Diamond, 71.07 on SWE-bench Verified and 52.3 on OSWorld-Verified. But it also lands at 17.78 on tau3-bench banking, which is a reminder that broad agent models still have sharp edges. NVIDIA says Super VL beats Nemotron 3 Ultra on Hermes Bench and LangChain evals, despite Ultra having about 4.6 times more total parameters.

There’s a second act here too. NVIDIA has shipped Nemotron 3.5 Bolt 74B-A9B for local agents, with native 2-bit quantization that fits into 24GB of memory on RTX and DGX Spark systems. A smaller Bolt E4B model is due on October 23 for voice assistants and embedded agents. The company clearly wants one family that can scale from a local PC to a data-center box without changing the pitch.