PSA
While APEX in general has previously shown itself to be rather reliable in everyday application, I doubt the mainline APEX repo allocations of attention precision in particular. GLM-4.7-Flash has rather lean attention tensors. For example, attn_kv_a_mqa, which is of paramount importance, would contibute 16515072 b at BF16 instead of 5308416 b at Q4_K. That's a 61 MB (total, not per layer!) diff in filesize. There are similar gains to be had in other attention tensors.
I'm providing the original APEX allocations in this repo, but an artifact with more conservative attention allocation is in the works. Stay tuned.
What's that?
APEX (Adaptive Precision for EXpert Models) quantizations of MuXodious/GLM-4.7-Flash-absolute-heresy using default configurations provided in the APEX repo as of 10/13/2026.
using imatrix by mradermacher found here
quantized with llama.cpp #1692f9e50
For roleplaying, I personally strongly recommend nohurry's SillyTavern presets for GLM 4.7 Flash models.
Available Files
| File | Profile | Size | Label |
|---|---|---|---|
| GLM-4.7-Flash-absolute-heresy-APEX-iBalanced.gguf | I-Balanced | 21 GB | Q5_K_M |
| GLM-4.7-Flash-absolute-heresy-APEX-iQuality.gguf | I-Quality | 18 GB | Q4_K_M |
| GLM-4.7-Flash-absolute-heresy-APEX-iCompact.gguf | I-Compact | 14 GB | Q3_K_M |
| GLM-4.7-Flash-absolute-heresy-APEX-iMini.gguf | I-Mini | 12 GB | IQ3_XXS |
Disclosure
My only contribution is compute. This is neither my abliteration artifact, nor my quantization methodology. WYSIWYG. Have fun.
Cheers
Z.ai - the base model.
MuXodious - for teaching the model to swear.
EmanuelOverride - for sparking my interest in local quantizations and APEX in particular.
mradermacher - for the imatrix + myriads of quants we all benefit from.