GLM-5.3-Flash was Ox Alpha. Two GGUF files, ready for llama.cpp.
By Stealth Models, who tracked it as Ox Alpha before Z.ai revealed it.
StealthMark results for GLM-5.3-Flash (Ox Alpha):
Compare it with Qwen3.8-27B, the best local model in our tests
Files as text| File | Size | Pick it for |
|---|---|---|
Q5_K_XL/ |
240 GB in 6 parts | Near full quality, on machines with 256 GB or more of combined memory |
IQ2_XXS/ |
102 GB in 4 parts | The smallest size we recommend: 128 GB machines such as a 128 GB Mac, DGX Spark or Strix Halo, or a 32 GB GPU with 96 GB of RAM |
mmproj-F16.gguf |
1.1 GB | Image input, full precision |
mmproj-Q8_0.gguf |
0.62 GB | Image input, near lossless |
mmproj-Q6_K.gguf |
0.49 GB | Image input, smaller |
mmproj-Q4_K.gguf |
0.34 GB | Image input, smallest |
Run it
Download one size:
hf download stealth-model/GLM-5.3-Flash-GGUF --include "IQ2_XXS/*" "mmproj-Q8_0.gguf" --local-dir GLM-5.3-Flash
Point llama.cpp at the first part; it loads the rest:
llama-server -m GLM-5.3-Flash/IQ2_XXS/GLM-5.3-Flash-IQ2_XXS-00001-of-00004.gguf --mmproj GLM-5.3-Flash/mmproj-Q8_0.gguf --jinja
If the model is larger than your GPU, keep its expert layers in system RAM with --n-cpu-moe.
License
MIT, from Z.ai's GLM-5.3-Flash. The license file is included.
Ox Alpha card · StealthMark leaderboard · Follow us
Text and images: CC BY 4.0, stealthmodels.com.