GLM-5.3 is the largest open-weights model Z.ai has shipped: a 753B-parameter mixture-of-experts that activates 40B parameters per token. Its smallest official GGUF is 217 GB, so no MacBook can hold it, and the 4-bit MLX build that mlx-community documents for a single Mac Studio with 512 GB is 418 GB. The 93 GB Flash variant would fit a 128 GB Mac, but nothing can load it there yet. Below is the real file size for every quant, which Mac fits each one, and the single blocker that decides the Flash verdict.
Will GLM-5.3 run on your Mac? The short answer
Not on any MacBook. There is no 16 GB, 24 GB, 32 GB, 48 GB, 64 GB, 96 GB or 128 GB Mac that holds a single GLM-5.3 build, because the smallest quantized file is 217 GB. The floor is a 256 GB Mac Studio running the 1-bit or 2-bit quant, and even that comes with a real accuracy cost. The first Mac where the flagship is a comfortable daily driver is a 512 GB Mac Studio, and the specific build mlx-community ships for that machine is 418 GB.
GLM-5.3-Flash is the opposite story. Its smallest GGUF is 93 GB, which is small enough for a 128 GB Mac, but the weights declare an architecture that mainline llama.cpp does not know yet. Downloading it today gives you a file, not a working model. That distinction matters more than the size table, so it gets its own section below.
What GLM-5.3 actually is
Z.ai announced GLM-5.3 on August 14, 2026 and published the full weights on Hugging Face on August 28, 2026. The model reuses the GLM-5.2 base, so every capability gain comes from post-training rather than a new pre-training run (Unsloth).
The configuration published with the base repo is worth knowing, because it explains the memory profile (Z.ai):
- Architecture: glm_moe_dsa, a mixture-of-experts with dynamic sparse attention.
- 78 hidden layers, 256 routed experts, 8 experts active per token, 1 shared expert.
- Hidden size 6,144, MoE intermediate size 2,048.
- Context window: 1,048,576 tokens, with a vocabulary of 154,880.
- Reasoning is always on, with low, high and max effort levels.
The unquantized releases are not something a home lab touches. The default FP8 repository is 756 GB across 141 weight shards, and the BF16 sibling lands near 1.5 TB (Z.ai). Everything that follows is about the quantized builds, which are the only ones that fit a single Mac.
How many parameters does GLM-5.3 have? The number depends on who is counting
You will see 743B, 744B, 753B and 754B, and all four trace back to a real artifact. They count different things: the published family label, the base checkpoint, and each conversion's own metadata.
| Figure | Source | What it actually counts |
|---|---|---|
| 744B total, 40B active | Unsloth's GLM-5.3 guide (Unsloth) | The family label Z.ai and serving vendors publish |
| 753B | zai-org/GLM-5.3 and zai-org/GLM-5.3-BF16 safetensors metadata: 753,329,940,480 parameters (Z.ai) | The base checkpoint |
| 743B | mlx-community/GLM-5.3-4bit safetensors metadata: 743,377,019,904 parameters (mlx-community) | That one 4-bit conversion |
| 754B | unsloth/GLM-5.3-GGUF's own GGUF header: 753,864,139,008 parameters (Unsloth) | The GGUF conversion |
So the honest range is 743B to 754B, with 744B total and 40B active as the label to use in conversation. None of these numbers change the memory math: the smallest file is 217 GB either way.
The license is not MIT
GLM-5.2 shipped MIT. GLM-5.3 does not. The flagship carries a custom "GLM-5.3 License" from Z.AI, and the Hugging Face card reports the license as other rather than MIT (Z.ai).
Read the file and the change is smaller than the reaction suggests. The grant covers use, copying, modification, distribution, sublicensing, sale, fine-tuning and derivative works, with the usual requirement to keep the notice. The one real addition is a Model-as-a-Service clause: if you or your affiliates run a Model-as-a-Service business and your combined revenue exceeds 10 billion US dollars over any consecutive 12 months, you must pass Z.AI's security review before any commercial use (Z.AI). Routing requests to someone else's hosted model is carved out, and so are end-user products where model capability is embedded in a feature.
For a home lab, a researcher or a normal company, nothing changes. For a hyperscaler reselling inference, it does. GLM-5.3-Flash stayed MIT, so if license terms decide your deployment, Flash is the safer pick for reasons that have nothing to do with memory.
GLM-5.3 GGUF sizes, quant by quant
These are the real totals, recalculated by adding every shard of each Unsloth Dynamic 3.0 quant in the repository's file listing rather than copied from a summary table (Unsloth). The memory band column is the requirement Unsloth publishes for that quant class (Unsloth). The accuracy column is Unsloth's own top-1 measurement for each quant, rounded to a whole percent.
| Quant | Real file size | Unsloth memory band | Top-1 accuracy | Verdict |
|---|---|---|---|---|
| UD-IQ1_S | 217 GB | 1-bit: 223 GB | about 73% | Smallest build that exists. Fits a 256 GB Mac Studio. Heavy accuracy loss. |
| UD-IQ1_M | 228 GB | 1-bit: 223 GB | about 76% | The best 1-bit option. Unsloth's own round number for this class is 76%. |
| UD-IQ2_M | 239 GB | 2-bit: 245 GB | about 79% | Unsloth's recommendation for 256 GB devices. |
| UD-Q2_K_XL | 254 GB | 2-bit: 245 GB | about 81% | Biggest of the 2-bit class. Needs 256 GB with headroom to spare. |
| UD-IQ3_XXS | 282 GB | 3-bit: 290-360 GB | about 84% | Leaves the 256 GB tier behind. |
| UD-Q3_K_XL | 343 GB | 3-bit: 290-360 GB | about 89% | The first build with an accuracy story worth telling. |
| UD-IQ4_XS | 365 GB | 4-bit: 372-475 GB | about 91% | Sits just under the published 4-bit floor. |
| UD-Q4_K_XL | 467 GB | 4-bit: 372-475 GB | not published | The size that is cut off on the model card. Verified from the file listing. |
| UD-Q5_K_XL | 562 GB | 6-bit: 570 GB | not published | No Mac holds it. |
| UD-Q6_K_XL | 684 GB | 6-bit: 570 GB | not published | No Mac holds it. |
| Q8_0 | 801 GB | 8-bit: 810 GB | not published | Datacenter territory. |
| BF16 | 1,508 GB | 8-bit: 810 GB | not published | Datacenter territory. |
One note on that table. The Unsloth model card lists sizes only up to UD-IQ4_XS, so the UD-Q4_K_XL figure of 467 GB has to be read directly off the file listing in the repository, which is what we did here. If you went looking for the 4-bit GGUF size and found the card silent, that is why.
The RAM ladder Unsloth publishes
Unsloth states the requirement in whole-memory terms, counting RAM and VRAM together, because a mixture-of-experts model at these sizes spills across both. Their published bands are:
| Quant class | Total memory needed |
|---|---|
| 1-bit | 223 GB |
| 2-bit | 245 GB |
| 3-bit | 290-360 GB |
| 4-bit | 372-475 GB |
| 6-bit | 570 GB |
| 8-bit | 810 GB |
The same guide says the 2-bit UD-IQ2_M quant "uses 239GB of disk space [and] works well on 256GB RAM devices like a 2x NVIDIA DGX Sparks or a Mac Studio", and that the 1-bit quant will fit on 223 GB of RAM (Unsloth). Those two lines are the whole reason a 256 GB Mac Studio is the floor rather than a 512 GB machine.
Which Mac holds which GLM-5.3 build
This is the table to read before you download anything. The constraint is unified memory, and the file has to fit alongside the KV cache, the sparse attention indexer and the operating system, so the practical requirement sits above the raw file size, especially at long context.
| Mac unified memory | GLM-5.3 flagship builds that fit | Verdict |
|---|---|---|
| 16 GB | none | No. Smallest build is 217 GB. |
| 24 GB | none | No. |
| 32 GB | none | No. |
| 48 GB | none | No. Use GLM-5.3-Flash instead, once it loads. |
| 64 GB | none | No. |
| 96 GB | none | No. The closest thing is a 120B-class MoE model. |
| 128 GB | none | No. Still 89 GB short of the smallest quant. |
| 256 GB Mac Studio | UD-IQ1_S at 217 GB, UD-IQ2_M at 239 GB | The floor. 1-bit and 2-bit only, with a real accuracy cost. |
| 512 GB Mac Studio | everything up to UD-Q4_K_XL at 467 GB, plus the 418 GB MLX 4-bit build | The only Mac where GLM-5.3 is a genuine daily option. |
Note the shape of that table. The gap between 128 GB and 256 GB is where GLM-5.3 stops being a cloud problem and becomes a hardware purchase, and the gap between 256 GB and 512 GB is where it stops being a compromise. If your Mac has less than 256 GB of unified memory, no quant of this model runs on it, and no amount of tuning changes that.
Can a 512 GB Mac Studio actually run it?
Yes, and this is documented rather than inferred. The mlx-community 4-bit conversion of GLM-5.3 is 418 GB on disk across 91 shards, and its model card states plainly that it was made for people using a single Apple Mac Studio M3 Ultra with 512 GB (mlx-community).
Two details from that card are worth carrying forward. It is converted from the BF16 base rather than the FP8 release, and the author notes that other quantization recipes for this model start to overthink and redo decisions, while the standard 4-bit quant stays stable. So on a 512 GB machine the practical picks are the 418 GB MLX build for MLX, or the 467 GB UD-Q4_K_XL GGUF if you want to stay in the llama.cpp ecosystem.
Speed is a separate question and it depends on active parameters, not total size. A 40B-active mixture-of-experts decodes far faster than a dense 40B model, but at this memory footprint you are still bound by unified memory bandwidth, and a file that occupies most of your RAM leaves little room for a long context window. Plan the quant around the context length you actually need: at 1M tokens the KV cache and the attention indexer are a real line item, not a rounding error.
Why GLM-5.3-Flash will not load yet
GLM-5.3-Flash is the model most people should want. It is a 320B mixture-of-experts with 18B active parameters, MIT licensed, with a 1M-token context window and native image and video input (Z.ai). It is built on a new base rather than the GLM-5.2 one, with a hybrid of sparse and linear attention and manifold-constrained hyper-connections.
The sizes are friendly:
| Flash quant | Real file size |
|---|---|
| UD-IQ1_S | 93 GB |
| UD-IQ1_M | 98 GB |
| UD-IQ2_XXS | 102 GB |
| UD-Q2_K_XL | 109 GB |
| UD-IQ3_XXS | 120 GB |
| UD-Q3_K_XL | 148 GB |
| UD-IQ4_XS | 157 GB |
| UD-Q4_K_XL | 200 GB |
| UD-Q5_K_XL | 240 GB |
| Q8_0 | 341 GB |
| BF16 | 642 GB |
Those totals are recomputed from the shard listing in the Unsloth repository (Unsloth). At 93 GB for the smallest build, a 128 GB Mac has room for the weights, the cache and a working desktop.
The blocker is the architecture declaration, and it is visible in the file itself. The GLM-5.3-Flash GGUF metadata reports its architecture as glm5next. Mainline llama.cpp knows chatglm, glm4, glm4moe and glm-dsa, but not glm5next (llama.cpp). The pull request that adds it, PR 27773 "add GLM-5.3-Flash (GLM5-Next) support", was opened on August 26, 2026, last updated on September 12, 2026, and is still open and unmerged as of September 17, 2026 (llama.cpp). Three sibling pull requests target the same architecture and are also open.
That absent architecture constant cascades. Ollama lists both glm-5.3 and glm-5.3-flash, but in each case the only tag available is the cloud entry, so there is no local download to pull (Ollama). LM Studio has model pages for both and offers them through LM Studio Cloud rather than as local downloads (LM Studio). Neither is a surprise: both lean on llama.cpp for local GGUF inference, and llama.cpp cannot build a graph for an architecture it does not define.
Worth contrasting with the flagship, because the two differ in exactly the way that matters. The GLM-5.3 GGUF metadata declares glm-dsa, and glm-dsa is in the mainline architecture table today (Unsloth). So the giant model is the one that loads in a stock build, and the small one is the one that does not. Size and support are unrelated.
When PR 27773 merges, the Flash verdict flips from cannot load to runs, and 93 GB becomes a genuinely viable local model on a 128 GB Mac. Until then, treat the Flash GGUF as a download to watch, not a model to run.
What to run on your Mac instead
If the answer above is no, the useful next step is to pick a model that matches your memory rather than wait for a 754B-class model to shrink. A few anchors, with the full ladders in the linked posts:
| Your unified memory | What actually runs well | Where to look |
|---|---|---|
| 16 GB | 7B to 14B class quantized models | Best LLM for 16GB Mac |
| 24-48 GB | 20B to 30B class, including gpt-oss-20b at 24 GB | DeepSeek V4 VRAM requirements |
| 96 GB | GPT-OSS 120B and Qwen 3.5 122B-A10B, both at 96 GB | Kimi K3 VRAM requirements |
| 256 GB and up | 400B-class mixtures-of-experts, and the GLM-5.3 1-bit and 2-bit quants | Mac Studio M5 Ultra 512GB |
There is also a strand of work aimed at the opposite problem: running a 744B-class model on tiny memory by streaming weights from disk instead of holding them in RAM. It is a real technique with a brutal speed ceiling, and the honest math is worth reading before you plan around it: Run a 744B model on 25GB RAM.
To check a specific model against a specific Mac, the ModelFit engine answers the "will it fit" question directly, and the how much RAM for a local LLM guide covers the memory maths behind these tables.
FAQ
Will GLM-5.3 run on a 64 GB MacBook Pro?
No. The smallest GLM-5.3 build that exists is the 217 GB UD-IQ1_S quant, which is more than three times a 64 GB machine's total unified memory. No quant, offload setting or context trim closes that gap. GLM-5.3 needs at least 256 GB of unified memory, which means a Mac Studio rather than a laptop.
How much VRAM does GLM-5.3 need at 4-bit?
Two real answers. The MLX 4-bit build that mlx-community ships is 418 GB and is documented for a single 512 GB Mac Studio. The llama.cpp equivalent, UD-Q4_K_XL, is 467 GB across 11 shards. Unsloth's published memory band for the 4-bit class is 372-475 GB of combined RAM and VRAM, so both sit inside a 512 GB machine and neither fits anything smaller.
Can a 256 GB Mac Studio run GLM-5.3?
Yes, but only on the smallest quants. UD-IQ1_S is 217 GB and UD-IQ2_M is 239 GB, and Unsloth explicitly names 256 GB devices as the target for the 2-bit build. Expect a substantial accuracy cost: Unsloth measures top-1 accuracy around 73% for the 1-bit class and around 81% for the 2-bit class, against roughly 91% for a 4-bit build.
Is GLM-5.3 open source under MIT?
No. The flagship ships under a custom GLM-5.3 License from Z.AI. The grant is broad, covering use, modification, distribution and fine-tuning, but a Model-as-a-Service business above 10 billion US dollars in annual revenue needs Z.AI's security review first. GLM-5.3-Flash is the MIT-licensed model in the family.
Why does GLM-5.3-Flash need a llama.cpp pull request?
Because its GGUF files declare the architecture glm5next, and mainline llama.cpp does not define that architecture in its model table. Without it, llama.cpp cannot build an inference graph for the weights. PR 27773 adds the support and was still open and unmerged on September 17, 2026, so any Flash GGUF you download today will not load in a stock build. Ollama and LM Studio are affected for the same reason.
Sources
- Z.ai GLM-5.3 model card and config: https://huggingface.co/zai-org/GLM-5.3
- Z.ai GLM-5.3-Flash model card: https://huggingface.co/zai-org/GLM-5.3-Flash
- Unsloth GLM-5.3 GGUF repository and file listing: https://huggingface.co/unsloth/GLM-5.3-GGUF
- Unsloth GLM-5.3-Flash GGUF repository and file listing: https://huggingface.co/unsloth/GLM-5.3-Flash-GGUF
- Unsloth GLM-5.3 run guide, memory bands and accuracy table: https://unsloth.ai/docs/models/glm-5.3
- mlx-community GLM-5.3-4bit: https://huggingface.co/mlx-community/GLM-5.3-4bit
- llama.cpp pull request 27773 for GLM5-Next support: https://github.com/ggml-org/llama.cpp/pull/27773
- llama.cpp architecture table: https://github.com/ggml-org/llama.cpp/blob/master/src/llama-arch.cpp
- GLM-5.3 License text: https://huggingface.co/unsloth/GLM-5.3-GGUF/blob/main/LICENSE
- Ollama library entries: https://ollama.com/library/glm-5.3-flash
- LM Studio model page: https://lmstudio.ai/models/glm-5.3-flash
Match this model to a machine that can run it: by RAM tier for Apple Silicon, or by VRAM for an NVIDIA GPU.
The weekly local-AI refresh
New open-weight models, real Apple Silicon benchmarks, and the one model worth running on your Mac this week. Free, one email a week, unsubscribe anytime.
By subscribing you agree to our Privacy Policy and to receive the weekly email. Unsubscribe anytime.
Have questions? Reach out on X/Twitter