NVIDIA Nemotron 3.5 Super VL: open 120B-A12B vision-language model

NVIDIA released Nemotron 3.5 Super VL on October 15, 2026. The Hugging Face model cards give that release date, and NVIDIA's blog announced the model the same day. It is an open vision-language model with 120B total and 12B active parameters. The model card calls it a Mamba-2–Transformer hybrid mixture of experts (MoE) with multi-token prediction (MTP). It has 512 routed experts per MoE layer and a C-RADIO vision encoder. Context goes up to 1M tokens, but the card says individual deployment recipes may use a smaller window. Input is text, images and video (mp4, up to 2 minutes). Output is text. The licence is the OpenMDW License Agreement, version 1.1. NVIDIA's Hugging Face org has public BF16, NVFP4 and Base BF16 repos, plus an NVFP4-LiLiCorr speculative drafter. The org listing has no NVIDIA GGUF of this model. The BF16 card lists two deployment configurations, both on vLLM or SGLang: 4× H100 (TP4) and 2× GB200 (TP2). It names NVIDIA Hopper and Blackwell as the hardware and Linux as the operating system. The NVFP4 card lists 1× DGX Spark, 1× H200, 2× H100 or 1× GB200, and adds TensorRT-LLM as a runtime. Its DGX Spark rows list text and image input. Video appears only in the vLLM rows for H100, H200 and GB200. The card recommends the LiLiCorr drafter for all DGX Spark deployments. For local inference on DGX Spark, RTX Spark and RTX Pro 6000, it links to llama.cpp (a ggml-org collection) and Unsloth. The NVIDIA blog says the model ships with open weights, open datasets and training recipes. It also says the model runs on DGX Spark, DGX Station, RTX PRO and RTX systems, and on Jetson at the edge.