GGUF quantizations of AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored, an
abliteration of Qwen/Qwen3.6-27B. Validated to retain
the abliteration (0/100 refusals) and the base model's gsm8k capability across
the full ship list. Quantized from the BF16 source via llama.cpp with
imatrix calibration.
This is a text-only GGUF — the multimodal vision tower from the base is
not included. Abliteration affects refusal behavior on text inputs only;
the vision tower would otherwise be unchanged from upstream Qwen, and shipping
it adds ~3 GB per quant for no abliteration-related value. If multimodal
GGUF support is requested, an mmproj companion will be published separately.
Inheritance from the base model
AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored is an abliterated derivative
of Qwen/Qwen3.6-27B (Apache-2.0). The author reports KL divergence of
0.000492 from the base model — among the cleanest published Qwen3.6
abliterations to date. The abliteration is applied via weight projection
(no fine-tuning), so the model retains its training distribution; only
refusal-elliciting directions in activation space are projected out.
Validated downstream: an FP8 quant of this model scores 88.0% on gsm8k strict
(1319-question full set) versus the vanilla Qwen/Qwen3.6-27B-FP8 baseline
at 84.7% — abliteration removed refusals without measurable capability loss.
Quant size guide
Quant
File
Size
BPW
Audience
Q8_0
Qwen3.6-27B-AEON-Ultimate-Uncensored-Q8_0.gguf
26.6 GB
8.50
Reference. ≈BF16 by every metric. The "FP8-equivalent" build for users who want maximum quality.
Imatrix Q4. Note: atypically came in slightly worse than Q4_K_S on this model — see eval table. Pick Q4_K_S over IQ4_XS for AEON-7 specifically.
Q3_K_M
...-Q3_K_M.gguf
12.4 GB
3.95
12 GB cards (RTX 3060/4070 base).
IQ3_M
...-IQ3_M.gguf
11.7 GB
3.74
Imatrix Q3. Beats Q3_K_M on quality at lower BPW. Recommended over Q3_K_M for 12 GB cards.
Q2_K
...-Q2_K.gguf
10.0 GB
3.18
8–10 GB cards. Real perplexity hit (+7.6%) but capability and abliteration both intact in our eval.
Quality measurements
All numbers are computed on the BF16 source as the reference baseline.
PPL is on wikitext-2 test (100 chunks of 512 tokens). KLD is computed
against BF16 logits over the same chunks. Lower is better for both.
Quant
PPL
PPL/BF16
Mean KLD
Median KLD
99% KLD
(BF16)
7.184
1.0000
—
—
—
Q8_0
7.185
1.00014
0.0050
0.0006
0.015
Q6_K
7.211
1.00387
0.0057
0.0013
0.033
Q5_K_M
7.194
1.00144
0.0156
0.0031
0.101
Q4_K_M
7.237
1.00745
0.0281
0.0068
0.218
Q4_K_S
7.221
1.00524
0.0317
0.0080
0.273
IQ4_XS
7.290
1.01486
0.0298
0.0080
0.262
Q3_K_M
7.431
1.03442
0.0712
0.0241
0.717
IQ3_M
7.360
1.02448
0.0796
0.0307
0.819
Q2_K
7.730
1.07609
0.1710
0.0690
1.712
Behavioral evals (boundary quants)
The three boundary quants (highest, default, lowest) were tested directly:
Quant
Refusals (mlabonne100)
gsm8k strict
gsm8k flex
FP8 (vLLM, source-of-truth on full 1319-q gsm8k)
0/100
88.0%
89.5%
Q8_0 (50-q gsm8k slice)
0/100
88.0%
92.0%
Q4_K_M (50-q gsm8k slice)
0/100
84.0%
88.0%
Q2_K (50-q gsm8k slice)
0/100
90.0%
92.0%
Notes on the gsm8k 50-q slice: standard error at p=0.85 with n=50 is ~5pp.
Differences between Q8_0/Q4_K_M/Q2_K within ~10pp of each other are consistent
with sampling noise, not capability ordering. The PPL/KLD table above
captures the actual quality ordering. The important result is that
all three boundary quants retained 0/100 refusals, confirming the
abliteration survives even Q2_K's aggressive ~3.18 BPW.
The intermediate quants (Q6_K, Q5_K_M, Q4_K_S, IQ4_XS, Q3_K_M, IQ3_M)
were not directly tested for refusal/capability. PPL+KLD strictly bracketed
between the tested boundary quants, so we infer they fall within the same
behavioral envelope.
Speed (NVIDIA RTX A6000, full GPU offload, llama-bench)
Quant
pp512 (tok/s)
tg128 (tok/s)
Q8_0
1379
23.1
Q6_K
1169
27.8
Q5_K_M
1239
31.5
Q4_K_M
1207
35.4
Q4_K_S
1288
37.4
IQ4_XS
1368
38.7
Q3_K_M
1184
33.1
IQ3_M
1254
40.1
Q2_K
1036
40.8
Generation speed scales with quant size (memory-bandwidth-bound). Q8_0 → Q2_K
is +78% throughput. Prompt processing is roughly flat across quants
(compute-bound, not memory-bound).
These numbers are A6000-specific. Consumer cards (4080/4090, 24 GB) will
have different absolute throughput but similar relative ordering.
Search for kasimat/Qwen3.6-27B-AEON-Ultimate-Uncensored-GGUF in the LM Studio
model browser and pick a quant. The chat template is embedded in each GGUF.
Disabling thinking (Qwen3.x default-on)
Qwen3.x defaults to a <think>...</think> reasoning preamble. For most
inference and especially for benchmarking, disable it by passing
enable_thinking: false via the chat template:
python
1# Python OpenAI client against llama-server with --jinja:2client.chat.completions.create(3 model="aeon",4 messages=[...],5 extra_body={"chat_template_kwargs":{"enable_thinking":False}},6)
This is required to reproduce our eval numbers — thinking-on otherwise eats
the response budget on long prompts.
Quantization method
Source: BF16 weights from AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored
(~52 GB). NOT requantized from any FP8/INT8 intermediate; quants are
computed directly from the BF16 source for maximum precision.
Tool:llama.cpp HEAD (commit fc2b005, April 2026).
Built with CUDA 12.8.
Imatrix calibration: Bartowski's calibration_datav3.txt (Dampf-on-top-of-Kalomaze
v3, ~280 KB mixed English/code/multilingual). Computed against the BF16
source with --n-gpu-layers 55 partial offload (BF16 27B doesn't fit a
single 48 GB card fully). 200-chunk run, all 129 chunks of the calibration
corpus consumed. Final BF16 PPL on the calibration corpus = 6.93.
Quantization recipe: standard llama-quantize <bf16> <out> <quant>
with --imatrix for all quants except Q8_0 (where imatrix gives essentially
zero benefit).
Architecture: Qwen3.5 hybrid attention + Gated DeltaNet SSM. llama.cpp
registers this as MODEL_ARCH.QWEN35. The text-only language model is
produced via convert_hf_to_gguf.py's Qwen3_5TextModel handler.
Reproduction gotcha: BPE pre-tokenizer
If you re-run convert_hf_to_gguf.py from a fresh llama.cpp clone, you
will hit:
NotImplementedError: BPE pre-tokenizer was not recognized
chkhsh: 1444df51289cfa8063b96f0e62b1125440111bc79a52003ea14b6eac7016fd5f
AEON-7's tokenizer hash isn't registered upstream (the abliteration retraining
shifted the vocab layout from stock Qwen3.5). The fix is to add this block to
get_vocab_base_pre() in convert_hf_to_gguf.py, just after the existing
qwen35 entry:
python
1if chkhsh =="1444df51289cfa8063b96f0e62b1125440111bc79a52003ea14b6eac7016fd5f":2# ref: https://huggingface.co/AEON-7/Qwen3.6-27B-AEON-Ultimate-Uncensored3 res ="qwen35"
The pre-tokenizer behavior is structurally identical to stock Qwen3.5
(Sequence: Split-with-canonical-regex + ByteLevel); only the vocab differs.
Qwen3.6-27B-AEON-Ultimate-Uncensored.imatrix — the importance matrix used
to produce the imatrix-aware quants. Ship for reproducibility.
chat_template.jinja — the Qwen3.x chat template embedded in each GGUF;
also provided standalone for clients that don't read it from the GGUF.
Modelfile.example — Ollama Modelfile template pointing at the Q4_K_M.
Intended use & safety
This is an abliterated ("uncensored") model: the safety-tuning's refusal
behavior has been suppressed via weight-space projection. It will produce
content the upstream Qwen3.6-27B would refuse, including content that may
be harmful, illegal, or distressing. Use cases include:
Research on alignment, refusal mechanisms, and steering
Creative writing with adult / dark themes
Red-teaming scenarios
Tool use where overly-cautious refusals are themselves a safety hazard
(e.g. a medical-information assistant)
This model is not suitable for direct deployment to consumer products
without an additional safety layer between user input and model output.
The abliteration is intentional and load-bearing; do not try to "fix" it
with system prompts.
