Downloads · 30 days
8
19% of all-time downloads
Zynerji/Ektome-Qwen3.6-27B-PristinelyUncensored-NF4
Ektome-Qwen3.6-27B-PristinelyUncensored-NF4 is a machine learning model from Zynerji. Use it for the machine learning task on the model card, and read the license before you ship it in a product. The card lists the license as apache-2.0.
bitsandbytes 4-bit build of the bf16 parent. Vision tower retained. Zero training — the parent is a pure weight excision; this repo only quantises it.
Downloads · 30 days
8
19% of all-time downloads
All-time downloads
43
Public
Parameters
27.4B
17.9 GB on disk
Likes
0
Public
Click a slice to open those files.
.safetensors17.9 GB · 100%
How the weights are stored.
U824.8B · 91%
From the Hugging Face model README
bitsandbytes 4-bit build of the bf16 parent. Vision tower retained. Zero training — the parent is a pure weight excision; this repo only quantises it.
| bf16 parent | this (NF4) | |
|---|---|---|
| compliance (n=100) | 0.990 | 1.000 |
| MMLU-val (n=400) | 0.855 | 0.853 |
| foreign / degen / instr | 0.000 / 0.000 / 0.400 | 0.000 / 0.000 / 0.400 |
| size | ~51 GB | ~17 GB |
4-bit did not restore the refusal direction — compliance is unchanged within noise (the 0.990 → 1.000 move is one prompt in a hundred). The MMLU cost is 0.002, notably smaller than the 0.005 the 9B paid and the 0.025 a smaller build paid; larger models absorb 4-bit better.
GPTQ is impossible on this architecture — see the parent card for the Marlin/DeltaNet detail.
Ektomē: a rank-1 projected (orthogonalised against the harmless mean), norm-preserving
excision of the refusal direction from residual-write matrices (o_proj, out_proj,
down_proj). Zero training, zero fine-tuning, zero distillation — the delta is a linear
weight edit derived from activations. A configuration ships only if compliance clears target
and capability holds (dcap ≥ −SE) and generation quality holds (no code-switching,
degeneration, or instruction-following loss).