GLM-5.3-Flash-NVFP4

GLM-5.3-Flash-NVFP4

Gives you NVIDIA's NVFP4 quantized copy of ZAI's GLM-5.3-Flash, a multimodal model for reasoning, coding and agent tasks. Meant for developers who deploy on Blackwell GPUs with vLLM or SGLang.

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 quantized version of GLM-5.3-Flash, a language model from ZAI that reads text, images and video and answers in text. NVIDIA compressed the weights to the NVFP4 format with its Model Optimizer so the model is cheaper to serve.

The base model is a Mixture-of-Experts design for reasoning, coding and agent work. It has 320B parameters in total and 18B active at a time, with a context length up to 1M.

Who it suits: It suits developers and infrastructure teams who want a ready-made quantized model to deploy, not people who want a small file to try on a laptop.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Serve the model behind an agent system | Run multi-step coding tasks
AI platforms | ML engineers | Host a quantized model on Blackwell GPUs | Serve chat traffic with vLLM
Enterprise tools | RAG builders | Answer questions over documents with a long context | Read a large report in one pass
Chat products | Chatbot teams | Add a reasoning model that also reads images | Answer questions about a screenshot

If you do not have NVIDIA Blackwell hardware or a multi-GPU server, you can skip it, since the weights come to about 321 GB.

Where to use it: The card lists the runtimes, the hardware and the operating system it was prepared for.

• vLLM as the runtime engine
• SGLang as the runtime engine
• NVIDIA Blackwell GPUs
• Linux as the preferred operating system

Use case | Who | What they get back
Agent systems | Developers | Text answers from reasoning and tool use
Chatbots | Product teams | Chat replies that can take image or video input
RAG systems | Search and data teams | Answers grounded in long documents

How to set it up: You need a Linux machine with NVIDIA Blackwell GPUs, enough disk for roughly 321 GB of weights, and Python to fetch the files.

• Python with pip
• Disk space for the Safetensors weights
• vLLM or SGLang as the serving engine

1. Install the Hugging Face hub package.
2. Download the repository to your machine.
3. Load the folder in vLLM or SGLang on your GPUs.

pip install -U huggingface_hub
hf download nvidia/GLM-5.3-Flash-NVFP4

How to use it: Once a server is running, you send it text, images or video and read the text it returns. The card gives the full run steps for the supported engines, so follow those for your engine.

Input | What goes in
Text | A string prompt
Image | RGB image data
Video | MP4 or WebM files

Example: You send the running model a screenshot of a failing test and ask it to explain the error and suggest a fix. It answers with text that describes what went wrong and what to change.

For newcomers and experts: If you are new, know that this is a large model meant for servers. A first small step is to read the base model page and check your GPU against the Blackwell requirement before you download anything.

If you know the field, the model is a quantized Mixture-of-Experts with hybrid sparse and linear attention. NVIDIA calibrated the quantization on a news dataset and one of its own post-training chat datasets, and the card lists six benchmarks used for evaluation.

Download the NVFP4 weights
Load them in vLLM or SGLang
Send text, image or video
Get a text answer back

Good to know: NVIDIA says it did not develop the base model, which comes from ZAI. The card also advises testing with your own data before you deploy.

• License: mit
• Cost to run: none named, but it needs NVIDIA Blackwell GPU hardware
• Data it reaches: files you download, and whatever you send the model
• Not for: machines without Blackwell GPUs, since that is the hardware the card names

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: Model Optimizer
• Parameters: 321B
• Weights: about 321 GB in U8
• Formats: Safetensors
• Downloads in the last 30 days: 79,326
• Likes: 186
• License: mit
• Based on: zai-org/GLM-5.3-Flash
• Quantized versions: 6
• Fine-tunes: 1
• Tags: safetensors, glm5_next, nvidia, modelopt, glm-5, quantized, fp4, image-text-to-text, conversational, 8-bit

Links

• Model: https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4
• Files: https://huggingface.co/nvidia/GLM-5.3-Flash-NVFP4/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download nvidia/GLM-5.3-Flash-NVFP4

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