GLM 5.3 and GLM 5.3 Flash are live in LU Labs Cloud
Both new Z.ai models are in the catalog. Flash is on the Hosted plan, so every plan has it, and the flagship is on Pro and Max. Both think before every answer, which is why there is now an Effort button next to the Brain button.
Z.ai put the GLM 5.3 Flash weights on Hugging Face on 27 August 2026 and the GLM 5.3 flagship a day later. We spent that week reading the repositories rather than the announcement, and wrote up what we found. The uncomfortable part of that write-up: the smallest usable build of the flagship is 216.7 GB and the smallest build of Flash is 93.1 GB, which is a machine almost nobody owns.
Both models are now in LU Labs Cloud. GLM 5.3 Flash is on the Hosted plan, which means it is on every plan. GLM 5.3 is on Pro and Max. Refresh the model picker in the web app or the desktop app and they are there.
The new Effort button
Both of these are reasoning models, and both think before every answer. Every message gets a reasoning pass, hard question or not.
So this release brings a control that the catalog did not have before. Next to the Brain button there is now an Effort button, and it decides how much the model thinks before it answers. It is in the web app today. On the desktop app it arrives with the next update, 2.6.8, on Windows, Linux and Mac.
Most reasoning models offer three settings: Low, Medium and High. GLM 5.3 and GLM 5.3 Flash take a fourth one above those, Max, so the button shows four on those two and three everywhere else. It starts on High, which is what LU Labs has always sent when the Brain button was on, so nothing about your existing chats changes until you move it yourself.
That choice costs something. A higher setting produces more output tokens, and output tokens are what draw on your credits. Low is the saving: on the flagship it answers a simple question in about a fifth of the tokens High spends. Max is worth reaching for on a problem that earns it, and on the rest it mostly buys you a longer wait.
The button is not specific to GLM. It applies to every reasoning model in the catalog, so it changes how the models you already use behave too.
What each one is
GLM 5.3 is the 753.9 billion parameter flagship, text only, with a 1,048,576 token context window. It ships under Z.ai's own glm-5.3 licence rather than MIT, which matters if you were planning to run it yourself commercially and does not affect you at all if you use it here.
GLM 5.3 Flash is the smaller one at 320.8 billion parameters, of which 18 billion are active per token. It carries the same million token window, it is MIT licensed, and it reads images natively. We checked that last part against the upstream rather than the spec sheet: Flash takes an image and describes it correctly, and the flagship refuses image input outright.
Tools work natively on both, so Agent mode and the coding agent call functions the normal way with no prompt tricks in between.
| GLM 5.3 | GLM 5.3 Flash | |
|---|---|---|
| Plan | Pro and Max | every plan |
| Context | 1,048,576 tokens | 1,048,576 tokens |
| Thinks | always | always |
| Reads images | no | yes |
| Native tools | yes | yes |
| Licence | Z.ai glm-5.3 | MIT |
The practical split is the one the size suggests. Flash is the model to leave selected, since it sees images and a million tokens of context holds a whole repository or a book. The flagship is what you bring in when a problem is genuinely hard and the extra thinking is worth what it costs.
If you would rather run them yourself
Both sets of weights are public, so this is a real option if you have the hardware for it. We wrote the hardware side up in detail.
How to run GLM 5.3 on your own computer has the measured quant sizes for both models, the memory math on top of those file sizes, and the reversal that costs people an afternoon: the 754 billion parameter flagship runs in a stock llama.cpp build, while the much smaller Flash still does not. As of 2 September 2026, pull request 27754 for Flash is open. Unsloth's GGUF packs for Flash have been up since 26 August, and MLX builds of the flagship arrived on 2 September as mlx-community/GLM-5.3-4bit.
GLM 5.3 vs GLM 5.2 is the field by field diff of the two config files. Every architecture value matches and both report the same parameter count down to the last digit, so the flagship is GLM 5.2's body with different post-training. That single fact explains the llama.cpp result above, and it also explains why the licence got stricter while the hardware requirements did not move.
Also on the API
If you use LU Cloud through the OpenAI compatible endpoint, both models are there under zai-org/GLM-5.3 and zai-org/GLM-5.3-Flash. Point any OpenAI style client at https://lu-labs.ai/api/inference/v1 and they appear in the model list with everything else.
Not on Cloud yet? See the plans. Every tier includes chat, code, image and video in one studio, and upgrading buys volume rather than access to a hidden shelf.