About
weighted/imatrix quants of a Heretic merge of https://huggingface.co/Cyclone-Labs/Eclipsed-Phoenix-26B-A4B
Gemma 4 26B-A4B (Heretic trial 93: 21/100 refusals, KL 0.0128). Text GGUFs only — no mmproj.
These are i1 (importance-matrix) GGUFs in the mradermacher naming style, converted with llama.cpp on an NVIDIA GB10. Calibration: froggeric/imatrix groups_merged.txt (llama-imatrix -c 512).
Abliteration: p-e-w/heretic.
Usage
If you are unsure how to use GGUF files, refer to one of TheBloke's READMEs for more details, including on how to concatenate multi-part files.
Provided Quants
(sorted by size, not necessarily quality. IQ-quants are often preferable over similar sized non-IQ quants)
| Link | Type | Size/GB | Notes |
|---|---|---|---|
| GGUF | imatrix | 0.1 | imatrix file (for creating your own quants) |
| GGUF | i1-IQ1_S | 8.3 | for the desperate |
| GGUF | i1-IQ1_M | 8.7 | mostly desperate |
| GGUF | i1-IQ2_XXS | 9.3 | |
| GGUF | i1-IQ2_XS | 9.8 | |
| GGUF | i1-IQ2_S | 9.9 | |
| GGUF | i1-IQ2_M | 10.4 | |
| GGUF | i1-Q2_K | 10.6 | IQ3_XXS probably better |
| GGUF | i1-Q2_K_S | 10.6 | very low quality |
| GGUF | i1-IQ3_XXS | 11.3 | lower quality |
| GGUF | i1-IQ3_XS | 11.6 | |
| GGUF | i1-IQ3_S | 12.2 | beats Q3_K* |
| GGUF | i1-Q3_K_S | 12.2 | IQ3_XS probably better |
| GGUF | i1-IQ3_M | 12.4 | |
| GGUF | i1-Q3_K_M | 13.3 | IQ3_S probably better |
| GGUF | i1-Q3_K_L | 13.8 | IQ3_M probably better |
| GGUF | i1-IQ4_XS | 13.9 | |
| GGUF | i1-Q4_0 | 14.5 | fast, low quality |
| GGUF | i1-Q4_K_S | 15.5 | optimal size/speed/quality |
| GGUF | i1-Q4_1 | 16.0 | |
| GGUF | i1-Q4_K_M | 16.8 | fast, recommended |
| GGUF | i1-Q5_K_S | 18.0 | |
| GGUF | i1-Q5_K_M | 19.1 | |
| GGUF | i1-Q6_K | 22.6 | practically like static Q6_K |
Here is a handy graph by ikawrakow comparing some lower-quality quant types (lower is better):
And here are Artefact2's thoughts on the matter: https://gist.github.com/Artefact2/b5f810600771265fc1e39442288e8ec9
Thanks
Source weights by Cyclone-Labs. Heretic by p-e-w. Conversion with llama.cpp.