EDGE OS
Z.ai released GLM-5.3-Flash on August 26, 2026 as a faster, lighter sibling to its GLM-5.3 flagship. It is the first natively multimodal model in the GLM-5 series, built to read screenshots, documents and video as well as text, and it is published with open weights under the MIT license.
This guide covers what GLM 5.3 Flash is, how it is built, what it is good at and the hardware you need to run it on-premise.
What is GLM 5.3 Flash?
GLM 5.3 Flash is a mixture-of-experts model. It holds 320 billion parameters, but a router sends each token to only 8 of its 288 experts, so just 18 billion parameters do the work on any given token. That keeps it fast while giving it the knowledge of a much larger model.
- Parameters: 320B total, 18B active per token
- Experts: 288 experts, 8 active per token
- Attention: layers alternate between linear attention and sparse attention, which Z.ai says cuts attention compute by about 3× and the KV cache by about 4.4× compared with GLM-5.3
- Input and output: text, images and video in; text out
- Context window: up to 1 million tokens according to Z.ai and NVIDIA; some hosts set lower limits, so check the endpoint you use
- License: MIT
- Released formats: FP8 (native) and NVFP4 (about 181 GB)
What is GLM 5.3 Flash good at?
Z.ai positions GLM 5.3 Flash for work that mixes vision and action:
- Turning screenshots and designs into working UI code
- Computer use and GUI automation
- Creating and editing office documents: slides, PDFs, Word and Excel files
- Document research and financial workflows
Z.ai reports a score of 63.4 on DeepSWE v1.1, up from 46.2 for GLM-5.2. Independent tester Artificial Analysis gave it 57 on its Intelligence Index, ranking it 3rd of 109 models it measured.
How much VRAM does GLM 5.3 Flash need?
All 320B parameters must sit in GPU memory, even though only 18B are active per token. Here is what that means at each precision:
| Precision | Weights | 1 user, 8K context | 16 users, 8K context |
|---|---|---|---|
| FP8 | 320 GB | 337 GB | 342 GB |
| NVFP4 (4-bit) | 181 GB | 192 GB | 196 GB |
| BF16 | 640 GB | 673 GB | 678 GB |
Thanks to its hybrid attention, each user’s KV cache stays small, so adding users costs little extra memory compared with dense models.
How to run GLM 5.3 Flash on-premise
- GPU memory: about 337 GB at FP8 or 192 GB at NVFP4 for one user, and about 342 GB for a team of 16 at FP8
- Supported GPUs: NVIDIA lists B200, H200 and H20 GPUs for GLM-5.3-Flash. NVFP4 runs natively on Blackwell GPUs such as B200
- Software: vLLM, the runtime NVIDIA uses for its GLM-5.3-Flash service
- The model: the weights from Hugging Face (zai-org/GLM-5.3-Flash), about 320 GB at FP8
For a team of 16 at FP8, the smallest setups that fit are: NVIDIA DGX B200 (Fits on 1 system); AMD Instinct MI300X (2 GPUs); NVIDIA L40S (8 GPUs).
Our GLM 5.3 Flash hardware requirements page has VRAM for every precision, user count and context length, plus fine-tuning needs. Or adjust the numbers yourself in the LLM VRAM calculator.
Running GLM 5.3 Flash privately with Vanik Edge
Vanik Edge OS runs open-weight models like GLM 5.3 Flash on NVIDIA DGX, AMD Instinct and other GPU servers inside your building. Screenshots, documents and video your team shares with the model stay on your network.
Frequently asked questions
What is GLM 5.3 Flash?
GLM-5.3-Flash is a 320B-parameter mixture-of-experts model from Z.ai with 18B parameters active per token. Released on August 26, 2026, it reads text, images and video and is aimed at coding, computer use and document work.
How big is GLM 5.3 Flash?
GLM 5.3 Flash has 320 billion parameters in total, with 18 billion working on each token. The native FP8 weights take about 320 GB and the official NVFP4 build about 181 GB.
Is GLM 5.3 Flash open source?
Yes. Z.ai released the weights on Hugging Face (zai-org/GLM-5.3-Flash) under the MIT license, so you can self-host it and use it commercially.
Does GLM 5.3 Flash have vision?
Yes. It accepts text, images and video (mp4, mov and webm) as input and produces text.
How much VRAM do I need to run GLM 5.3 Flash?
About 337 GB at FP8 for one user with 8K context, or about 192 GB with the NVFP4 build. Serving a team of 16 at FP8 needs about 342 GB.
Can I run GLM 5.3 Flash locally?
Yes, on data center GPUs. Even at NVFP4 it needs about 192 GB, more than consumer graphics cards hold. A team of 16 at FP8 fits on NVIDIA DGX B200, AMD Instinct MI300X (2 GPUs).
Run open-weight models inside your building
Tell us the models and team size you have in mind. We will size the GPU server with you and set up Vanik Edge OS on it.