How to Run GLM-5.3 on Your Own Computer
Flash on 27 August, the flagship on 28 August. Here is what the 754B flagship and the 321B Flash actually need in disk and memory, which one your machine can hold, and the surprise: the big one runs in stock llama.cpp and the small one does not.
Z.ai shipped GLM-5.3 to their API before they shipped the weights, and then did something unusual: they kept the weights back for a while. For that stretch the model was something you could rent but not hold. The company said the delay was for safety evaluation and hardening, and was open about the reason. The same capabilities that help defenders find bugs early help attackers once anyone can download the file.
Z.ai has not published a precise launch date for that API in any primary source we could find. What the sources do support is the order and the month: the API was live first, and both it and the weights were in the field by late August 2026.
The weights themselves 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 at 14:21 UTC, and nobody builds a GGUF from weights that do not exist. Either Flash was already public on the 26th or Unsloth had early access, so treat 26 August as the lower bound.
This post is the honest hardware math. What each one is, what it costs in disk and memory, and the one thing that will waste your afternoon if nobody tells you first.
Two models wearing the same name
| GLM-5.3 | GLM-5.3-Flash | |
|---|---|---|
| Parameters | 753.9 billion | 320.8 billion |
| Active per token | not published | 18 billion |
| Context | 1,048,576 tokens | 1,048,576 tokens |
| Original weights | 756 GB | 328 GB |
| Licence | custom glm-5.3 | MIT |
| Reads images | no | yes, natively |
Both ship at FP8 out of the box, which is already half the size of the usual FP16 release, and both carry a one million token context window. That is not a marketing round number, it is max_position_embeddings: 1048576 sitting in the config file.
Flash is MIT, which is about as permissive as software gets. You can use it commercially, change it, ship it inside a product, and nobody has to be asked. The flagship is not MIT. It carries Z.ai's own glm-5.3 licence, listed as other with the licence name glm-5.3 and a LICENSE file in the repository. Commercial use is allowed, with one condition worth knowing about: if you run a model-as-a-service business and you and your affiliates turn over more than 10 billion US dollars in any twelve consecutive months, you need to pass a safety review by Z.ai before using it commercially. That threshold excludes almost everyone reading this, but it is the sort of clause you want to have read rather than heard about.
One number we are deliberately not giving you: how many parameters the flagship activates per token. It is a mixture-of-experts model with 256 experts, 8 of them picked per token plus 1 that always runs, across 78 layers. Z.ai has not published the resulting active count, so neither will we. If you see a confident figure somewhere, check whether it is quietly GLM-5.2's number.
What it costs on disk
Nobody runs raw FP8 weights at home. You run a quantised build, and the community had those up within hours. Below are the actual sizes of the Unsloth packs, read on the evening of 28 August 2026.
GLM-5.3, the flagship
| Quant | Size | Realistically for |
|---|---|---|
| UD-IQ1_S | 216.7 GB | a 256 GB machine, and it will be tight |
| UD-IQ1_M | 228.5 GB | 256 GB with less room to spare |
| UD-Q2_K_XL | 253.9 GB | 384 GB and up |
| UD-Q3_K_XL | 343.0 GB | 512 GB |
| UD-Q4_K_XL | 467.3 GB | a server |
GLM-5.3-Flash
| Quant | Size | Realistically for |
|---|---|---|
| UD-IQ1_S | 93.1 GB | 128 GB unified memory |
| UD-IQ1_M | 97.6 GB | 128 GB |
| UD-Q2_K_XL | 108.7 GB | 128 GB, getting tight |
| UD-IQ3_XXS | 120.4 GB | 128 GB only with a short context |
| UD-Q3_K_XL | 147.5 GB | 192 GB and up |
| UD-IQ4_XS | 156.8 GB | 192 GB |
| UD-Q4_K_XL | 199.7 GB | 256 GB |
These are file sizes, not memory requirements. The weights are the floor, not the ceiling. Your context window, the key value cache and the runtime all want their own space on top, and on a million token model that cache is not a rounding error. A build that fits your disk with 10 GB to spare does not fit your memory.
If you want Flash to look at images, there is a second file: mmproj-F16.gguf, 1.13 GB. Download it alongside the model. Without it Flash is a very capable text model that quietly cannot see, and the failure mode is not an error message, it is just a model that never mentions your picture.
The part that will waste your afternoon
The 754 billion parameter flagship runs in a normal, current llama.cpp. The smaller Flash does not, which is the wrong way round from what anyone would guess.
The flagship writes itself into GGUF as architecture glm-dsa, and LLM_ARCH_GLM_DSA is in the llama.cpp main branch today. Pull, build, run.
Flash writes itself as glm5next, and that architecture is not in the main branch. Unsloth's own model card points at llama.cpp PR 27754 or their desktop app instead. So the model most people can actually fit is the one that needs a pull request to run, and the model almost nobody can fit is the one that just works.
We checked the main branch again on 2 September 2026, and the situation has not moved. It carries chatglm, glm4, glm4moe and glm-dsa, and no glm5next. PR 27754 is still open, last touched on 1 September 2026 at 21:53 UTC. This will change, probably soon, quite possibly before you read this. Check before you spend an hour rebuilding. If your llama.cpp greets a Flash download with an unknown architecture error, that is this, and it is not something you misconfigured.
Flash has had GGUF builds from Unsloth since 26 August 2026, so the files are not the bottleneck. The loader is.
One setting worth knowing before you start
The flagship takes a reasoning_effort parameter with three settings: low, high and max.
If you do not pass it, you get max. If you pass something that is not one of those three words, you also get max. There is no middle default and no warning, so a typo in that field costs you a long, expensive think on every single message. On a model this size that is the difference between a usable assistant and one you stop opening.
Start at low and raise it when a task actually needs it.
So which one should you get
If you have 128 GB of unified memory, Flash at UD-IQ1_S or UD-IQ1_M is the only one of the two that fits, and it is MIT licensed, reads images, and is the one Z.ai positions as the cheap workhorse. Budget for a llama.cpp built from the pull request, or use a runtime that already ships support.
If you have 256 GB or more, the flagship at UD-IQ1_S becomes possible and runs on a stock build. Whether a 1-bit quant of a 754B model beats a 2-bit quant of a 321B one is a real question with no honest answer yet, because nobody has had these files long enough to say.
If you have a normal computer, neither of these fits, and no amount of quantisation is going to change that this year. It is a 756 GB model. Run something built for your hardware and use a hosted endpoint when you want the big one.
Worth working out which of the three you are in before the download starts rather than after.
On a Mac, there is now an MLX route
As of 2 September 2026 the flagship has MLX conversions at mlx-community/GLM-5.3-4bit and mlx-community/GLM-5.3-mixed-4_5bit, both created that day. Neither has meaningful download numbers yet, so nobody has reported back on how they behave. If you have a Mac large enough to be in the conversation at all, they are worth knowing about, and worth treating as new.
What we have not verified
Z.ai publishes benchmark numbers for GLM-5.3, including a jump on CyberGym from 77.2 percent for GLM-5.2 to 84.5 percent, and on ExploitBench from 24.4 to 54.4 percent. Those are the vendor's own figures from the vendor's own model card, measured with their harness and their settings.
We have not reproduced them. Nobody outside Z.ai has, yet. Treat them the way you would treat any number a company publishes about its own product on launch day: interesting, directionally useful, and not evidence.
When we have run these models ourselves for long enough to say something you cannot get from a press release, we will write that post instead.