Qwen3.6-35B-A3B-NVFP4

Qwen3.6-35B-A3B-NVFP4

A pre-quantized NVFP4 copy of Alibaba's Qwen3.6-35B-A3B, made by NVIDIA with Model Optimizer. It takes text, images and video, and runs on vLLM with Hopper or Blackwell GPUs.

Made by nvidia on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: This is NVIDIA's NVFP4 quantization of Alibaba's Qwen3.6-35B-A3B language model, made with the Model Optimizer tool. It is a mixture-of-experts model with hybrid attention, 35B parameters in total and 3B activated. The weights take about 18.7 GB. The card notes that the original model was not developed by NVIDIA.

Who it suits: Developers who want a ready-made quantized model to serve on NVIDIA GPUs, and teams who build chatbots, agents or document question systems on their own hardware.

Industry | Who uses it | How they use it | Example
Software | Developers | Serve a pre-quantized model behind an app | A chatbot backend on vLLM
Agents | Agent builders | Power an agent system with a model they host | A tool-using assistant
Search | RAG builders | Answer questions over documents with their own model | A support knowledge bot
Research | ML researchers | Study NVFP4 quantization against the original model | A side-by-side quality check

If you have no NVIDIA Hopper or Blackwell GPU, you can skip it, since the card lists only those two hardware families.

Where to use it:

• vLLM, the only runtime engine the card names
• Linux machines with NVIDIA Hopper or Blackwell GPUs
• Windows, macOS or Linux with Python, for downloading the files

Use case | Who | What they get back
Text chat | Developers | Text answers to written prompts
Image questions | Analysts | Text answers about an image
Video questions | Media teams | Text answers about an MP4 or WebM video

How to set it up:

• Python, to run the Hugging Face download tool
• Disk space for about 18.7 GB of weights
• A Linux machine with an NVIDIA Hopper or Blackwell GPU and vLLM

1. Install the Hugging Face hub tools with pip.
2. Download the model files with the hf command.
3. Serve the model with vLLM, following the model card.

pip install -U huggingface_hub
hf download nvidia/Qwen3.6-35B-A3B-NVFP4

How to use it: Serve the model with vLLM, then send it text, images or video from your own app. The card lists text as the only output and a context length of up to 262K.

Example: You run the model in vLLM on a Hopper GPU and send it a long report with a question about one section. It returns a text answer.

For newcomers and experts: Start with a short text prompt before you add images or video. The card advises testing the model on your own data before you rely on it.

The card says NVIDIA quantized the model with Model Optimizer version 0.44.0 and calls this release NVFP4 version 1.0. Calibration used a news-article set and a conversation set. The card lists evaluation sets such as MMLU Pro and GPQA Diamond, but the text shown gives no scores.

Download the model files
Serve it with vLLM on a Hopper or Blackwell GPU
Send text, images or video
Text answers come back

Good to know:

• License: Apache 2.0
• Cost to run: none named, but it needs an NVIDIA Hopper or Blackwell GPU
• Data it reaches: downloads files from Hugging Face, then runs on your own machine
• Not for: CPU-only machines, since the card says it is built for NVIDIA GPU systems
• Limits: the card marks the training data details as undisclosed

On Hugging Face, 13 November 2026

• Task: text generation
• Library: Model Optimizer
• Parameters: 18.7B
• Weights: about 18.7 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 4,104,345
• Likes: 678
• License: apache-2.0
• Based on: Qwen/Qwen3.6-35B-A3B
• Quantized versions: 10
• Tags: safetensors, qwen3_5_moe, nvidia, modelopt, qwen3.6, quantized, fp4, text-generation, conversational, 8-bit

Links

• Model: https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4
• Files: https://huggingface.co/nvidia/Qwen3.6-35B-A3B-NVFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/Qwen3.6-35B-A3B-NVFP4

Then follow the usage in its model card. Nothing in this kit runs by itself.