Mixed-precision (UVMAX) quantization of zai-org/GLM-5.3-BF16, converted with mlx-lm from the bf16 release.
UVMAX assigns bit widths per tensor class instead of quantizing uniformly.
| tensor class | precision | parameters | size | share |
|---|---|---|---|---|
| Expert FFN gate/up | 4-bit, group 128 | 483B | 239.1 GiB | 59.1% |
| Expert FFN down | 5-bit, group 128 | 242B | 147.7 GiB | 36.5% |
| Attention + DSA indexer + dense MLP | 8-bit, group 64 | 13.7B | 13.6 GiB | 3.4% |
| Shared experts | 8-bit, group 64 | 2.8B | 2.8 GiB | 0.7% |
| Embeddings, LM head | 4-bit, group 64 | 1.9B | 1.0 GiB | 0.2% |
| Routers | bf16 | 0.1B | 0.2 GiB | <0.1% |
| Norms | bf16 | — | — | <0.1% |
| total | 4.67 bits/weight | 743B | 404 GiB |
Teacher-forced against the bf16 release on identical tokens, 48 windows of
1025 tokens. KLD is KL(bf16 ‖ quant) over the full output distribution and
is corpus-specific.
UVMAX (4.67 bits/weight)
| corpus | ppl bf16 | ppl UVMAX | ratio | mean KLD | median KLD | top-1 agreement |
|---|---|---|---|---|---|---|
| Linux kernel C | 1.434 | 1.459 | 1.02× | 0.040 | 0.0002 | 96.8% |
| XNU kernel C | 2.435 | 2.472 | 1.02× | 0.056 | 0.0027 | 93.4% |
| JavaScriptCore C++ | 1.757 | 1.791 | 1.02× | 0.055 | 0.0007 | 95.1% |
| English prose | 2.687 | 2.750 | 1.02× | 0.063 | 0.0076 | 92.4% |
| all | 2.015 | 2.053 | 1.02× | 0.053 | 0.0012 | 94.4% |
Uniform 4-bit (4.50 bits/weight)
| corpus | ppl bf16 | ppl 4-bit | ratio | mean KLD | median KLD | top-1 agreement |
|---|---|---|---|---|---|---|
| Linux kernel C | 1.434 | 1.492 | 1.04× | 0.076 | 0.0004 | 95.3% |
| XNU kernel C | 2.435 | 2.522 | 1.04× | 0.109 | 0.0063 | 91.0% |
| JavaScriptCore C++ | 1.757 | 1.841 | 1.05× | 0.102 | 0.0014 | 93.4% |
| English prose | 2.687 | 2.909 | 1.08× | 0.143 | 0.0209 | 88.6% |
| all | 2.015 | 2.119 | 1.05× | 0.107 | 0.0028 | 92.1% |
Requires a recent mlx-lm with GLM-5.3 (glm_moe_dsa) support.
The model fits on a single 512 GB machine:
mlx_lm.server --model kernelpool/GLM-5.3-5bit-UVMAX
Sampling follows the base model: temperature 1.0, top-p 0.95.
model-00001-of-00113.safetensors 5.4 GBmodel-00002-of-00113.safetensors 4.1 GBmodel-00023-of-00113.safetensors 4.1 GBmodel-00071-of-00113.safetensors 4.1 GBmodel-00029-of-00113.safetensors 4.1 GBmodel-00035-of-00113.safetensors 4.1 GBmodel-00041-of-00113.safetensors 4.1 GBmodel-00047-of-00113.safetensors 4.1 GBmodel-00053-of-00113.safetensors 4.1 GBmodel-00059-of-00113.safetensors 4.1 GBmodel-00065-of-00113.safetensors 4.1 GBmodel-00095-of-00113.safetensors 4.1 GBmodel-00107-of-00113.safetensors 4.1 GBmodel-00083-of-00113.safetensors 4.1 GBmodel-00089-of-00113.safetensors 4.1 GBmodel-00077-of-00113.safetensors 4.1 GBmodel-00011-of-00113.safetensors 4.1 GBmodel-00017-of-00113.safetensors 4.1 GBmodel-00005-of-00113.safetensors 4.1 GBmodel-00101-of-00113.safetensors 4.1 GBmodel-00038-of-00113.safetensors 4.0 GBmodel-00050-of-00113.safetensors 4.0 GBmodel-00020-of-00113.safetensors 4.0 GBmodel-00044-of-00113.safetensors 4.0 GBmodel-00098-of-00113.safetensors 4.0 GBmodel-00026-of-00113.safetensors 4.0 GBmodel-00110-of-00113.safetensors 4.0 GBmodel-00074-of-00113.safetensors 4.0 GBmodel-00080-of-00113.safetensors 4.0 GBmodel-00048-of-00113.safetensors 4.0 GBmodel-00087-of-00113.safetensors 4.0 GBmodel-00015-of-00113.safetensors 4.0 GBmodel-00062-of-00113.safetensors 4.0 GBmodel-00030-of-00113.safetensors 4.0 GBmodel-00108-of-00113.safetensors 4.0 GBmodel-00104-of-00113.safetensors 4.0 GBmodel-00090-of-00113.safetensors 4.0 GBmodel-00056-of-00113.safetensors 4.0 GBmodel-00014-of-00113.safetensors 4.0 GBmodel-00068-of-00113.safetensors 4.0 GBmodel-00027-of-00113.safetensors 4.0 GBmodel-00057-of-00113.safetensors 4.0 GBmodel-00081-of-00113.safetensors 4.0 GBmodel-00032-of-00113.safetensors 4.0 GBmodel-00072-of-00113.safetensors 4.0 GBmodel-00069-of-00113.safetensors 4.0 GBmodel-00086-of-00113.safetensors 4.0 GBmodel-00054-of-00113.safetensors 4.0 GBmodel-00063-of-00113.safetensors 4.0 GBmodel-00008-of-00113.safetensors 4.0 GBmodel-00021-of-00113.safetensors 4.0 GBmodel-00042-of-00113.safetensors 4.0 GBmodel-00093-of-00113.safetensors 4.0 GBmodel-00099-of-00113.safetensors 4.0 GBmodel-00084-of-00113.safetensors 4.0 GBmodel-00092-of-00113.safetensors 4.0 GBmodel-00060-of-00113.safetensors 4.0 GBmodel-00078-of-00113.safetensors 4.0 GBmodel-00024-of-00113.safetensors 4.0 GBmodel-00033-of-00113.safetensors 4.0 GBmodel-00039-of-00113.safetensors 4.0 GBmodel-00036-of-00113.safetensors 4.0 GBmodel-00111-of-00113.safetensors 4.0 GBmodel-00075-of-00113.safetensors 4.0 GBmodel-00105-of-00113.safetensors 4.0 GBmodel-00102-of-00113.safetensors 4.0 GBmodel-00018-of-00113.safetensors 4.0 GBmodel-00009-of-00113.safetensors 4.0 GBmodel-00066-of-00113.safetensors 4.0 GBmodel-00003-of-00113.safetensors 4.0 GBmodel-00006-of-00113.safetensors 4.0 GBmodel-00051-of-00113.safetensors 4.0 GBmodel-00045-of-00113.safetensors 4.0 GBmodel-00096-of-00113.safetensors 4.0 GBmodel-00012-of-00113.safetensors 4.0 GBmodel-00022-of-00113.safetensors 3.4 GBmodel-00040-of-00113.safetensors 3.4 GBmodel-00049-of-00113.safetensors 3.4 GBmodel-00064-of-00113.safetensors 3.4 GBmodel-00106-of-00113.safetensors 3.4 GBmodel-00112-of-00113.safetensors 3.4 GBmodel-00016-of-00113.safetensors 3.4 GBmodel-00037-of-00113.safetensors 3.4 GBmodel-00052-of-00113.safetensors 3.4 GBmodel-00067-of-00113.safetensors 3.4 GBmodel-00070-of-00113.safetensors 3.4 GBmodel-00076-of-00113.safetensors 3.4 GBmodel-00082-of-00113.safetensors 3.4 GBmodel-00085-of-00113.safetensors 3.4 GBmodel-00091-of-00113.safetensors 3.4 GBmodel-00094-of-00113.safetensors 3.4 GBmodel-00013-of-00113.safetensors 3.4 GBmodel-00025-of-00113.safetensors 3.4 GBmodel-00034-of-00113.safetensors 3.4 GBmodel-00043-of-00113.safetensors 3.4 GBmodel-00046-of-00113.safetensors 3.4 GBmodel-00055-of-00113.safetensors 3.4 GBmodel-00061-of-00113.safetensors 3.4 GBmodel-00088-of-00113.safetensors 3.4 GBmodel-00097-of-00113.safetensors 3.4 GBmodel-00100-of-00113.safetensors 3.4 GBmodel-00058-of-00113.safetensors 3.4 GBmodel-00079-of-00113.safetensors 3.4 GBmodel-00103-of-00113.safetensors 3.4 GBmodel-00109-of-00113.safetensors 3.4 GBmodel-00019-of-00113.safetensors 3.4 GBmodel-00028-of-00113.safetensors 3.4 GBmodel-00031-of-00113.safetensors 3.4 GBmodel-00073-of-00113.safetensors 3.4 GBmodel-00004-of-00113.safetensors 3.4 GBmodel-00007-of-00113.safetensors 3.4 GBmodel-00010-of-00113.safetensors 3.4 GBmodel-00113-of-00113.safetensors 2.7 GBHave the files? Submit a torrent.
Submit a torrentHave the complete files and permission to share them? Replace the example folder path. The script checks the files, makes a torrent and submits its magnet and checksums. Keep your torrent client seeding. Get your community key from your settings, and read the script before running it.
curl -fsSL https://pirateface.co/package.sh | bash -s -- --repo kernelpool/GLM-5.3-5bit-UVMAX ./model-folderzai-org/GLM-5.3-BF16