Nemotron 3.5 Super VL
NVIDIA's open 120B mixture-of-experts multimodal agent model with 12B active parameters, released with weights, data and recipes.
Overview
NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, an open multimodal mixture-of-experts model with roughly 120B total and 12B active parameters for agentic workflows that combine reasoning, tool use, and text, image, document and video understanding. NVIDIA reports 80.8% on CharXiv (84% with a Python interpreter), 63.44% on Hermes Bench and 87.05% on LangChain Evals, and claims up to 5x higher agentic throughput than open models of similar size. Base, post-trained and NVFP4 checkpoints ship with training recipes, data and RL environments.
This is a sourced directory entry. Features and capabilities are reported in the linked sources; this entry does not document independent Aymorphic testing.
Key features
- About 120B total / 12B active parameters; third-party coverage describes a Mamba-2/Transformer hybrid with a 1M-token context (not stated in NVIDIA's blog)
- CharXiv 80.8% (84% with code interpretation) versus 63% for Nemotron 3 Nano Omni, per NVIDIA
- Hermes Bench 63.44% and LangChain Evals 87.05%, both above Nemotron 3 Ultra per NVIDIA
- Base, post-trained and NVFP4-quantized checkpoints plus training data, recipes and RL environments
- Runs locally on NVIDIA DGX Spark; NVFP4 targets Blackwell GPUs
- Available on Hugging Face, ModelScope, OpenRouter and NVIDIA NIM, with vLLM, SGLang and Ollama support
Use cases
- Multimodal document and chart understanding inside agent pipelines
- Local agentic coding and tool use on DGX Spark
- Fine-tuning from open base checkpoints and recipes
Pricing
Open weights (license named OpenMDW-1.1 by Artificial Analysis; NVIDIA's blog does not name it). Hosted pricing on OpenRouter, NIM and Baseten is unknown.
Pricing reflects the source record and may change. Confirm current terms with the provider.
Dates and sources
- Released
- 15 Oct 2026
- Record updated
- 17 Oct 2026
- Record last checked
- 17 Oct 2026