GLM-5.3-Flash-BF16

GLM-5.3-Flash-BF16

A 320B-parameter multimodal model with 18B active that reads text and images, with an adjustable thinking budget and a hybrid attention design for long context. For teams with serious GPU infrastructure.

Made by zai-org 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: Within the GLM-5 series, GLM-5.3-Flash is the first model built to handle images and text together from the start. It has 320B total parameters, with only 18B active per token, and it reads both images and text. This repository holds the full BF16 weights, about 643 GB. The authors say it beats GLM-5.2 at a tenth of the price and approaches a leading proprietary model on coding and agent tests.

Who it suits: It suits teams with large GPU servers who want an open model for coding, agents and long documents that also reads images.

Industry | Who uses it | How they use it | Example
Software | Engineering teams | Run a large open model for coding and agent work | A coding assistant that works across a big repository
AI infrastructure | Serving engineers | Deploy it with a serving framework on a GPU cluster | Hosting the model for internal users
Enterprise | Applications teams | Process long documents with images and text | Reading a long report with charts
Research | Model researchers | Study a sparse and linear attention hybrid | Testing the thinking levels on a task set

You can skip it if you have a single consumer GPU, since the full weights are about 643 GB and the card points to server-class serving tools.

Where to use it:

• SGLang, vLLM, TokenSpeed, KTransformers or Unsloth for serving, which the card lists
• The Transformers library
• The vendor's hosted API platform, which the card names as an alternative to running it yourself
• Windows, macOS or Linux, with Python for the download

Use case | Who | What they get back
Coding and agent tasks | Developers | Code and multi-step actions from the model
Image plus text questions | Analysts | Answers that use both a picture and text
Long-context work | Researchers | Replies over long inputs at a lower serving cost
Adjust thinking effort | App builders | Low, high or max reasoning on request

How to set it up:

• A GPU server sized for a 320B model, since the BF16 weights are about 643 GB
• Python and pip for the download
• A serving framework or the Transformers library
• Fast storage for the huge files

1. Install the Hugging Face hub package with pip.
2. Download the repository files, which are huge.
3. Serve the model with one of the frameworks the card lists, following its own guide.

pip install -U huggingface_hub
hf download zai-org/GLM-5.3-Flash-BF16

How to use it: You serve the model with a supported framework and chat with it. Two settings matter, and the card says how to use them.

Tool | What it does
reasoning_effort | Sets the thinking budget to low, high or max, with max as the default
clear_thinking | Defaults to false, and the card says to pass true for chat use

Example: A team serves the model and sends a chart image with a question about it. They leave the thinking level at max for a hard question. They set clear_thinking to true because it is a chat setting. The model answers using both the image and the text.

For newcomers and experts: If you are new to this, this model needs server hardware, so try it first through the vendor's hosted API or a smaller model. Read the serving guide for your framework before downloading.

If you know the field, the card describes a hybrid of sparse and linear attention that cuts long-context serving cost, plus manifold-constrained hyper-connections for scaling efficiency. It was trained on a 30T-token multimodal corpus. For benchmark reproduction the card says to keep the default max thinking level.

Download the BF16 weights
Choose a serving framework
Set the thinking level
Send text or an image with a question
Read the answer

Good to know:

• Limits: the card names no minimum hardware, and the weights alone are about 643 GB.
• Claims: the price and benchmark comparisons come from the authors.
• Settings: reasoning_effort falls back to max for any value other than low or high.
• License: MIT.
• Cost to run: no price is named, and the card points to the vendor's hosted API as an alternative to your own servers.
• Data it reaches: it copies huge model files to your computer and reads the text and images you give it.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 321B
• Weights: about 643 GB in BF16; about 161 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 68,087
• Likes: 85
• License: mit
• Languages: en, zh
• Quantized versions: 41
• Fine-tunes: 5
• Tags: transformers, safetensors, glm5_next, image-text-to-text, conversational

Links

• Model: https://huggingface.co/zai-org/GLM-5.3-Flash-BF16
• Files: https://huggingface.co/zai-org/GLM-5.3-Flash-BF16/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download zai-org/GLM-5.3-Flash-BF16

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