NVIDIA Releases Multimodal Model Nemotron 3.5 Super VL

NVIDIA released the open multimodal model Nemotron 3.5 Super VL with 120 billion total and 12 billion active parameters, up to 1 million tokens of context, and support for mp4 videos up to 2 minutes, as well as the compact Bolt 74B-A9B 24GB, which fits in 24 GB of VRAM.

Compiled by Sergey KostenchukPublished 2026-10-16Updated 2026-10-17

2026-10-16 Research NVIDIA

🤖 NVIDIA Releases Multimodal Model Nemotron 3.5 Super VL

The open model works with text, images, and mp4 videos up to 2 minutes: 120 billion total and 12 billion active parameters, a Mamba-2 and transformer hybrid, 512 experts in each MoE layer, and up to 1 million tokens of context. The reasoning mode is toggled in the chat template.

🌍 Weights, datasets, and training recipes are open under the OpenMDW-1.1 license — companies can fine-tune the model on their own data and deploy it in-house, without relying on closed APIs. The focus on documents (OCR, tables, PDF) is aimed at agentic business processes.

👤 Super VL is already available on Hugging Face and OpenRouter, and the compact Bolt 74B-A9B with 2-bit quantization fits in 24 GB of VRAM: agentic tasks — web, terminal, code — can be run locally on RTX without the cloud.

Source 1: https://blogs.nvidia.com/blog/nemotron-3-5-super-vl/ Source 2: https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16

Sources