Run GLM-5.3 Locally: Real Quant Sizes, the llama.cpp Surprise, and the reasoning_effort Trap

Z.ai published the GLM-5.3-Flash weights on 27 August 2026 at 10:33 UTC and the GLM-5.3 flagship on 28 August 2026 at 15:22 UTC. The models were on Z.ai's own API first, though I could not find a primary source that dates that launch, so I am not going to put a day on it. I spent the evening reading the actual repositories instead of the announcement. Here is what is worth knowing before you start a download measured in hundreds of gigabytes.

TL;DR

• Two models: GLM-5.3 at 753.9B params, weights out 28 August, and GLM-5.3-Flash at 320.8B (18B active), weights out 27 August.
• Smallest usable builds: 216.7 GB flagship, 93.1 GB Flash.
• The flagship runs in stock llama.cpp. Flash does not yet. Yes, that way round.
• reasoning_effort defaults to max, including when you typo it.
• Flash is MIT. The flagship is not.

The architecture surprise

I expected the smaller model to have better support. The opposite is true, and the reason is interesting.

I diffed the two config files:

Field	GLM-5.2	GLM-5.3
architectures	GlmMoeDsaForCausalLM	GlmMoeDsaForCausalLM
num_hidden_layers	78	78
hidden_size	6144	6144
n_routed_experts	256	256
num_experts_per_tok	8	8
max_position_embeddings	1048576	1048576
Params (GGUF metadata)	753,864,139,008	753,864,139,008

Identical. Every field, and the parameter count down to the last digit. GLM-5.3 is GLM-5.2 with different post-training, exactly as Z.ai described it at the API launch.

So the flagship's GGUF architecture is glm-dsa, which llama.cpp has supported since GLM-5.2:

$ grep 'LLM_ARCH_GLM' src/llama-arch.cpp
{ LLM_ARCH_GLM4,     "glm4"     },
{ LLM_ARCH_GLM4_MOE, "glm4moe"  },
{ LLM_ARCH_GLM_DSA,  "glm-dsa"  },

Flash is the genuinely new architecture. It reports itself as glm5next, which is still not in the main branch on 2 September 2026. PR 27754 is open, and Unsloth's card points at it or at their desktop app.

If you get an unknown architecture error loading a Flash GGUF, that is this. Check whether the PR has landed before you spend an hour rebuilding, because this will change.

Actual sizes

Unsloth packs, read on the evening of 28 August 2026. The Flash pack is dated 26 August 2026 and the flagship pack 28 August, which is worth noting because a GGUF cannot predate the weights it was built from, so 26 August is the lower bound for Flash. MLX builds of the flagship went up on 2 September 2026 as mlx-community/GLM-5.3-4bit.

GLM-5.3

UD-IQ1_S     216.7 GB
UD-IQ1_M     228.5 GB
UD-Q2_K_XL   253.9 GB
UD-Q3_K_XL   343.0 GB
UD-Q4_K_XL   467.3 GB

GLM-5.3-Flash

UD-IQ1_S      93.1 GB
UD-IQ1_M      97.6 GB
UD-Q2_K_XL   108.7 GB
UD-IQ3_XXS   120.4 GB
UD-Q3_K_XL   147.5 GB
UD-IQ4_XS    156.8 GB
UD-Q4_K_XL   199.7 GB

These are file sizes. Your KV cache is extra, and on a model advertising 1,048,576 tokens of context that is not a rounding error.

Flash is natively multimodal, and the vision half is a separate 1.13 GB file:

mmproj-F16.gguf    1.13 GB
mmproj-BF16.gguf   1.16 GB

Skip it and Flash silently cannot see images. There is no warning.

Running it

Flash, per Unsloth's card, once the PR is in your build:

llama serve -hf unsloth/GLM-5.3-Flash-GGUF:UD-Q4_K_XL

The flagship works on a stock build today, if you have somewhere to put 216.7 GB.

The default that will bite you

GLM-5.3 takes reasoning_effort with three values: low, high, max.

From the model card: it defaults to max if not passed, or if set to any other value.

There is no middle default, and an invalid value does not error, it silently gives you the most expensive setting. So a typo in that field buys you a long think on every message. Start at low.

One caveat, because I measured it rather than read it. Through a hosted OpenAI-compatible provider the behaviour is not the same: on the same arithmetic prompt the flagship spent 4 output tokens at low, 10 at medium, 11 at high and 48 at max, and sending no field at all landed at 30, not at max. medium is honoured there as a real middle rung even though the card does not list it. So the model card describes the model, and whatever sits in front of it may have its own opinion. Measure your own stack before you trust either.

Licensing, which went backwards

• GLM-5.2: MIT
• GLM-5.3-Flash: MIT
• GLM-5.3 flagship: Z.ai's own glm-5.3 licence, not MIT

Same architecture and same parameter count as the MIT-licensed 5.2, different terms. If you are shipping something commercial, that is a real decision and not a formality.

Z.ai was open about the reasoning: they held the weights back for safety evaluation, and named the concern. Their own numbers put ExploitBench at 54.4 percent versus 24.4 for GLM-5.2, and CyberGym at 84.5 versus 77.2. Those are vendor figures from a vendor harness, unreproduced by anyone outside the company at the time of writing, and the delay makes more sense once you see which benchmark moved most.

If the hardware is the blocker

Both of these are now in the LU Labs Cloud catalog, which is the product I work on, so read this section with that in mind. GLM-5.3-Flash sits on the Hosted plan, so it is on every plan, and the GLM-5.3 flagship is on Pro and Max. The catalog is served to the clients rather than compiled into them, so they appear in the model picker in the web app and in the desktop app without anyone installing an update.

Both models think before every answer, so there is now an Effort button next to the Brain button, and it controls how much the model thinks. It is in the web app today; on the desktop it arrives with the next update. Most reasoning models in the catalog offer three settings, Low, Medium and High, and these two add a fourth above them, Max. It starts on High. A higher setting means more output tokens and therefore more credits, so Low is the saving. Tools run natively on both, and Flash takes images.

What I did not check

The benchmarks. I have not run them and neither has anyone else independently yet. Everything above about sizes, parameters, architectures and licences comes from the HuggingFace API, the config files and the llama.cpp source, all of which answer in seconds and none of which are a press release.

If you are picking a model to actually run this week and you do not have 128 GB of memory, none of this is your model. That is a legitimate answer too.