NVIDIA
The official handle for NVIDIA.
5.0•35 reviews•2.3K followers
The official handle for NVIDIA.
5.0•35 reviews•2.3K followers
NVIDIA, inventor of the GPU, which creates interactive graphics on laptops, workstations, mobile devices, notebooks, PCs, and more. We created the world’s largest gaming platform and the world’s fastest supercomputer. We are the brains of self-driving cars, intelligent machines, and IoT.
This is the 33rd launch from NVIDIA. View more
Nemotron 3.5 Super VL
Launching today
Nemotron 3.5 Super VL is NVIDIA’s open, customizable multimodal model for specialized agents. It is a 120B mixture-of-experts model with 12B active parameters that handles text, image, document, and video input, plus reasoning and tool use, in one model. NVIDIA claims up to 5x higher agentic throughput on a single GPU than open models its size. Checkpoints, training recipes, and datasets are on Hugging Face, with a NIM container.
Free
Launch tags:Artificial Intelligence
Launch Team
Dropbox SignSimple, secure eSignatures for the way your team works.
Promoted
Nemotron 3.5 Super VL is NVIDIA’s open, customizable multimodal model for specialized agent workflows. It combines reasoning, instruction following, tool use, and text, image, document and video understanding in one model.
Problem: Agents that deal with documents, charts and video often need a separate text model and a separate vision-language model, plus the glue between them. That adds cost, latency and moving parts.
Solution: One 120B mixture-of-experts model with 12B active parameters per token. It takes text, image, document and video input, and also reasons and calls tools. It’s distilled from Nemotron 3 Ultra, per NVIDIA.
What makes it different: it’s open and built to be customized. NVIDIA releases base, post-trained and NVFP4-quantized checkpoints, plus training recipes, data and RL environments. Hugging Face also hosts 45 open vision-language datasets. You can fine-tune it with NVIDIA NeMo, using SFT, LoRA or reinforcement learning.
Key features and claims:
Up to 5x higher throughput for agentic tasks on a single GPU than other open models in its size class, with over 600 output tokens per second on a Baseten endpoint
80.8% on CharXiv chart reasoning, and 84% with Python code interpretation
4x faster completion of 10,000 coding tasks on NVIDIA B200 GPUs
Leads Gemma-4-31B-it and Sarvam-105B on five Indic benchmarks, after 10B Indic tokens were added in training
With NVIDIA Dynamo on 16 B200 GPUs, it serves up to 8x as many users and cuts first-token latency by 85%
An independent EXO Labs evaluation on DGX Spark placed it on the Pareto frontier for accuracy and task completion time, as cited by NVIDIA
Where to run it: Hugging Face, NVIDIA NIM, NVIDIA Dynamo and ModelScope. It’s also on hosted providers such as Baseten, Together AI, Fireworks AI, DeepInfra and CoreWeave. You can run it with vLLM, SGLang, Ollama and LM Studio, and it’s on Amazon SageMaker JumpStart, Google Cloud, Microsoft Foundry and OCI.
Pricing: the model is free to download and use under the OpenMDW 1.1 license, including commercially. Running it needs your own GPUs or a hosted provider.
Who it’s for: developers building specialized agents that read documents, charts, images or video and want an open model they can customize.
Download on Hugging Face · Read the technical report · Customization cookbook · Read the announcement