NVIDIA Nemotron 3.5 Super VL

Nemotron 3.5 Super VL, released October 15, 2026 (announced October 13), is NVIDIA's flagship open-weight model in the Nemotron 3.5 family. It is a 120-billion-parameter mixture-of-experts model with roughly 12 billion active parameters per token, built on a hybrid Mamba-2/Transformer LatentMoE decoder with a C-RADIO v4-H vision encoder. It reads text, images, short videos, and PDFs, with a 1-million-token context window, and is positioned for multimodal agentic enterprise tasks.

Weights ship under the OpenMDW-1.1 license, which allows commercial use, fine-tuning, and redistribution. The model is on Hugging Face, NVIDIA NIM, OpenRouter, and ModelScope; NVIDIA published an image-based reinforcement-learning recipe in its NeMo-RL repository, and llama.cpp contributors had GGUF conversion paths ready almost immediately. NVIDIA names deployment targets of 1 DGX Spark, 1 H200, 2 H100, or 1 GB200.

NVIDIA claims a post-trained version of Super VL beats Claude Opus 5.5 and GPT-6.1 Sol on specific specialized task benchmarks. That is a narrower, company-reported claim, not an independent general-reasoning result — public trackers still rank closed models ahead on broad benchmarks.

Provider-reported (NVIDIA): post-trained Super VL beats Claude Opus 5.5 and GPT-6.1 Sol on specialized task benchmarks.

Independent: no general-reasoning leaderboard rerun found; public trackers still rank closed models ahead.