glm-5.3-flash
Architecture mixture of experts45 LayersLicense mitMaximum context 1,024kVocabulary 154,880
At Q4_K_M it is a 187.04 GiB file, and it fits on 1 of the 18 machines we measure (only one: Mac Studio with M5 Ultra and 512 GB). Here the file is what fills the memory: even at its maximum context of 1024k tokens, the context adds 22.00 GiB — and the same content in German takes ×1.49 the tokens of English.
Model size at Q4_K_M
Measured by Local AI Scope
Context memory at 8k tokens
Measured by Local AI Scope
What German costs versus English
Measured by Local AI Scope
Videos: glm-5.3-flash 3
The best walk-throughs we found. Press play and it opens right here.
The videos live on YouTube, and nothing loads from there until you press play (no-cookie mode). The thumbnail is served by us.
Sizes of the real published files
These are the sizes of the files actually published by the author, added up when the model ships split into parts. Not an estimate from the parameter count.
| Compression | Model size |
|---|---|
Q8_0 |
318.05 GiB |
Q6_K |
261.74 GiB |
Q5_K_M |
226.96 GiB |
Q5_K_S |
212.26 GiB |
Q4_K_M |
187.04 GiB |
Q4_K_S |
176.30 GiB |
IQ4_XS |
165.54 GiB |
Q3_K_M |
144.37 GiB |
This model also reads images. For that it needs a second file, the vision projector (mmproj-GLM-5.3-Flash-BF16-f16.gguf, 1.05 GiB), published next to the model and not included in the figures on this page: they count text only. If you are going to send it images, add it to the model size.
Context memory
Computed from the attention pattern this model declares layer by layer, not from the generic formula. That is why the figure is often far smaller than other calculators tell you. This model stores its context memory compressed (latent attention): per layer and token, one vector of 512 values instead of its 64 heads of 256. We count it as one head of 512 for the key and another for the value, which is a little more than it really takes (one latent serves both), so these figures err on the high side, never the low one.
- 8k tokens → 0.17 GiB
- 32k tokens → 0.69 GiB
- 128k tokens → 2.75 GiB
What it costs in each language
Measured by running this model’s own tokenizer over the same content in four languages. Tokens are what you pay for in context, in memory and in your cloud bill.
| Kind of text | EN | ES | FR | DE |
|---|---|---|---|---|
| Code | — | ×1.26 | ×1.41 | ×1.64 |
| Contracts | — | ×1.29 | ×1.36 | ×1.52 |
| Business email | — | ×1.22 | ×1.32 | ×1.32 |
| Support tickets | — | ×1.33 | ×1.51 | ×1.41 |
English is the baseline: every figure is how many times more tokens the same content costs in that language. Average across the whole corpus: DE ×1.49.
Where it fits
Fits means the model, its context memory and the working headroom all fit at the stated context, leaving the system its share. 1 of the 18 machine profiles tested (Q4_K_M, 8k tokens).
If you are willing to compress it more, it does run where the table says no: at Q3_K_M (144.37 GiB), up to 256k tokens of context, on PC with Ryzen AI Max+ PRO 495 and 192 GB unified. Below Q4_K_M the model loses a little quality, and at Q3_K_M you can tell; that is why the table above sticks to Q4_K_M.
Before you hit download: as of 6 October 2026, the way to run this model on one of these machines is llama.cpp: its stable v0.6.0, released on 5 October 2026, opens it, and so do its daily builds from b11279 onwards; earlier versions do not. Ollama and LM Studio offer it only in their cloud (there is no local download in their catalogues), and mlx-lm does not support its architecture. vLLM does support it, but with the original FP8 weights (305.81 GiB) in graphics-card memory, and no PC on our list has that much VRAM. One upside: the Q4_K_M file carries a 3.93 GiB layer for predicting several tokens at once that llama.cpp does not load, so with it the model needs 3.93 GiB less than the table counts.