GLM-5.3-Flash

GLM-5.3-Flash

zai-org's natively multimodal GLM-5 series model with 320B total and 18B active parameters, with a hybrid sparse and linear attention design and a thinking budget control. Suits teams with large GPU servers.

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: GLM-5.3-Flash is the first natively multimodal model in zai-org's GLM-5 series. It reads images and text and has 320B total parameters with 18B active. The card says it beats GLM-5.2 on benchmarks and real workloads. A hybrid of sparse and linear attention cuts the cost of serving long contexts, and Manifold-Constrained Hyper-Connections help it scale.

Who it suits: Teams with large GPU servers who want an open multimodal model for coding and agent work, and developers who prefer to call it through a hosted API.

Industry | Who uses it | How they use it | Example
Software | Agent developers | Run coding and agent workloads | Connect it to a coding agent
Research | Model evaluators | Test it on agent and coding benchmarks | Reproduce a benchmark with the default thinking level
Product teams | App builders | Add a multimodal model to a product | Send an image and a question

Skip it if you have a single computer, because the weights are about 321 GB in FP8.

Where to use it:

• Serving engines named in the card, including SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth
• The Z.ai API Platform, for hosted access
• Servers with enough GPU memory for 321 GB of weights

Use case | Who | What they get back
Coding and agent tasks | Developers | Code and tool steps from a long context model
Image and text questions | Product teams | Text answers that use both a picture and your words
Chat with a thinking budget | App builders | Replies at a low, high or max reasoning level

How to set it up:

• Python and the Hugging Face tools
• About 321 GB of storage for the FP8 weights
• One of the serving engines the card lists

1. Install the Hugging Face hub package.
2. Download the repository files.
3. Serve them with one of the listed engines, following that engine's own guide.

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

How to use it: Send chat requests to your served copy or to the hosted API. Two settings matter, listed below. The card text here shows no serving command.

Tool | What it does
reasoning_effort | Sets the thinking budget to low, high or max, and max is the default
clear_thinking | Controls whether earlier reasoning is cleared, and defaults to false
hf download | Copies the repository files to your computer

Example: Ask a coding question with reasoning_effort set to low. The model spends a smaller thinking budget and answers. For chat, pass clear_thinking as true, as the card advises.

For newcomers and experts: If you are new to this, use the hosted API for a first try, since running the weights yourself needs a big GPU server.

If you know the field, the card says to keep the default max level when you reproduce its benchmarks. Its footnotes name tests such as HLE with tools, NL2Repo, DeepSWE, Terminal-Bench 2.1, Toolathlon Verified, AutomationBench and BabyVision.

Images and text go in
Sparse and linear attention read the context
The thinking budget sets reasoning depth
Text comes out

Title: Parameters in billions
Total: 320 billion
Active: 18 billion

Good to know:

• The card supports English and Chinese.
• The reasoning_effort setting accepts low, high and max, and anything else counts as max.
• Hugging Face's own count shows 321B parameters, and the card says 320B.
• License: mit.
• Cost to run: none named for your own hardware. The hosted API is a separate service and its prices are not stated here.
• Data it reaches: your prompts and images. With the hosted API, they go to Z.ai.
• Not for: the source names none.

On Hugging Face, 13 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 321B
• Weights: about 321 GB in F8_E4M3; about 161 GB at 4-bit (estimate)
• Formats: Safetensors
• Downloads in the last 30 days: 6,397,774
• Likes: 3,138
• License: mit
• Languages: en, zh
• Quantized versions: 239
• Fine-tunes: 43
• Tags: transformers, safetensors, glm5_next, image-text-to-text, conversational, fp8

Links

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

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