Can You Run GLM-5.3 Locally? The Honest Hardware Math

Short answer: almost certainly not on the computer you own. The smallest usable build of the flagship is 216.7 GB, and the smallest build of the lighter GLM-5.3-Flash is 93.1 GB. Those are download sizes, before context and cache. If you have a 24 GB graphics card, this release is not for your machine, and no quant is going to change that.

That is the honest version, and the rest of this page is the detail behind it, checked against the actual files rather than the announcement.

What landed, and when

Z.ai shipped GLM-5.3 through their own API before they published the weights. We have not found a primary source that dates that API launch, so this page does not name a day for it. The downloads are documented. GLM-5.3-Flash went public on 27 August 2026 at 10:33 UTC, and the GLM-5.3 flagship followed on 28 August 2026 at 15:22 UTC. There is one wrinkle: Unsloth's GGUF pack for Flash is dated 26 August 2026, and nobody builds a GGUF from weights that do not exist, so treat 26 August as the lower bound for Flash.

The company said the delay was for safety evaluation and hardening, and named the reason: the model got substantially better at finding and exploiting software vulnerabilities, which helps defenders and attackers equally once anyone can download it.

Two models came out, and they are much less alike than the shared name suggests.

The licence difference matters more than most people will notice on launch day. Flash is MIT, which means commercial use, modification and redistribution without asking anyone. The flagship is not. It ships under Z.ai's own terms, so if you are building a product on it, that is a document to read rather than skim.

What it costs on disk

Nobody runs raw weights at home. You run a quantised build, and the community had those up within hours. Unsloth's GGUF pack for Flash is dated 26 August 2026 and the flagship pack 28 August, and MLX builds of the flagship arrived on 2 September 2026 as mlx-community/GLM-5.3-4bit. The sizes below are the Unsloth figures, read on the evening of 28 August 2026.

GLM-5.3, the flagship

GLM-5.3-Flash

For Flash there is a second file if you want it to look at images: mmproj-F16.gguf, 1.13 GB. Download it alongside the model. Leave it out and Flash is a strong text model that quietly cannot see, and there is no error message telling you so.

These are file sizes, not memory requirements. The weights are the floor. Your context window and its cache want their own space on top, and on a model advertising a million tokens that cache is not a rounding error. A build that fits your disk with room to spare does not automatically fit your memory.

The trap that reverses everything

The 754 billion parameter flagship runs in a normal, current llama.cpp. The smaller Flash does not. That costs people an afternoon when nobody warns them first.

That sounds backwards until you look at why. GLM-5.3 is architecturally identical to GLM-5.2: same 78 layers, same 256 experts, same million token window, and exactly the same parameter count down to the last digit. It is the same model body with different post-training. So its architecture, glm-dsa, has been supported in llama.cpp since GLM-5.2 and needs nothing new.

Flash is the genuinely new thing. It writes itself as glm5next, and that architecture is not in the llama.cpp main branch. Unsloth's model card points at pull request 27754 or their own desktop app instead.

We checked the main branch again on 2 September 2026: it carries chatglm, glm4, glm4moe and glm-dsa, and no glm5next. Pull request 27754 was still open that day. That will change, possibly before you read this. Check before you rebuild. If your llama.cpp meets a Flash download with an unknown architecture error, this is why, and you did not misconfigure anything.

One setting that decides whether it feels fast

GLM-5.3 takes a reasoning_effort parameter with three settings: low, high and max.

Leave it out and you get max. Misspell it and you also get max. There is no middle default and no warning. On a model this size that is a long, expensive think on every single message, and it is the most common reason a capable model feels unusable. Start at low and raise it when a task earns it.

Both models now run in LU Labs Cloud

If the numbers above rule your machine out, that is the ordinary result rather than a failure on your side, and there is a way around it. Both models are in the LU Labs Cloud catalog, and every plan and every credit pack reaches both of them. GLM-5.3-Flash draws 0.015 credits per input token and 0.05 per output token, the GLM-5.3 flagship 0.12 and 0.4, so a 5 EUR pack of 165,000 credits is roughly 412,000 output tokens on the flagship. They show up in the model picker in the web app and in the desktop app on Windows, Linux and Mac.

Because both models think before every answer, there is now an Effort button next to the Brain button, and it decides how much the model thinks before it replies. It is in the LU Labs Cloud 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, and GLM-5.3 and GLM-5.3-Flash add a fourth above those, Max, so the button shows four on those two and three everywhere else. It starts on High. A higher setting produces more output tokens, and output tokens are what draw on your credits, so Low is the saving and Max is worth reaching for on a problem that earns it. The button is not specific to GLM: it applies to every reasoning model in the catalog.

Tools run natively on both, so Agent mode and the coding agent work the normal way, and Flash takes images.

Is there an uncensored version

For Flash, yes, and legally so. The MIT licence permits finetunes and abliterations without permission, and community builds started appearing on Hugging Face the same day the weights did, including uncensored GGUF packs.

For the flagship the licence is Z.ai's own rather than MIT, so what you may publish is governed by that text rather than by convention. And at 216.7 GB for the smallest build, the practical limit arrives well before the legal one for almost everybody.

If uncensored local AI is what you are here for, the honest recommendation is to run a model built for consumer hardware. Our guide to running uncensored AI locally covers the ones that actually fit, and Qwen 3.8 27B is the current sweet spot for a single graphics card.

What we have not verified

Z.ai publishes benchmark gains for GLM-5.3 over GLM-5.2, including CyberGym rising from 77.2 percent to 84.5 percent and ExploitBench from 24.4 percent to 54.4 percent. Both figures are the vendor's own, from the vendor's harness, with the vendor's settings, published on launch day about the vendor's own product.

We have not reproduced them and, at the time of writing, nobody outside Z.ai has. Treat them as directionally interesting rather than as evidence. Every number about file sizes, parameters, architectures and licences on this page comes from the model repositories and the llama.cpp source, not from the announcement.

Locally Uncensored is AGPL-3.0 licensed and free to use. Built by PurpleDoubleD.