---

license: apache-2.0

license_link: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE

library_name: gguf

base_model: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF

base_model_relation: quantized

model_name: Qwen3.6-27B-Heretic-Cerebellum-GGUF

model_creator: Qwen

model_type: qwen3

quantized_by: deucebucket

pipeline_tag: image-text-to-text

tags:

- GGUF

- qwen3

- qwen

- quantized

- cerebellum

- imatrix

- hybrid-ssm

- mixed-precision

- 2-bit

- heretic

- uncensored

- abliterated

model-index:

• name: Qwen3.6-27B-Heretic-Cerebellum-GGUF

results:

- task:

name: Text Generation

type: text-generation

dataset:

name: AI2 Reasoning Challenge

type: ai2_arc

config: ARC-Challenge

split: test

metrics:

- name: normalized accuracy

type: acc_norm

value: 0.9693

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Qwen3.6-27B-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: HellaSwag

type: hellaswag

split: validation

metrics:

- name: accuracy

type: acc

value: 0.9014

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Qwen3.6-27B-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: MMLU-Redux

type: cais/mmlu

config: all

split: test

metrics:

- name: accuracy

type: acc

value: 0.7621

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Qwen3.6-27B-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: HumanEval+ (pass@1)

type: openai_humaneval

split: test

metrics:

- name: pass@1

type: pass@1

value: 0.8476

source:

name: Local audited benchmark run (RTX 3090, llama.cpp) — chat no-think harness

url: https://huggingface.co/deucebucket/Qwen3.6-27B-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

- task:

name: Text Generation

type: text-generation

dataset:

name: WikiText-2 Perplexity

type: wikitext

config: wikitext-2-raw-v1

split: test

metrics:

- name: perplexity

type: perplexity

value: 6.925

source:

name: Local audited benchmark run (RTX 3090, llama.cpp)

url: https://huggingface.co/deucebucket/Qwen3.6-27B-Heretic-Cerebellum-GGUF/tree/main/benchmark_results

---

Qwen 3.6 27B Heretic — Cerebellum GGUF

Sensitivity-guided mixed-precision quantization of

llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF,

which is itself a decensored variant of

Qwen/Qwen3.6-27B

produced by llmfan46 using Heretic v1.2.0.

All future Heretic versions of this build will live in this repository.

Version identifiers appear only in filenames, not in the repo name.

Files

| File | Size | Description |

|------|------|-------------|

| Qwen3.6-27B-Heretic-Cerebellum-v1-Q2_K_Mixed.gguf | 12.87 GB (12,866,587,680 bytes) | Cerebellum v4 recipe — recommended |

| Qwen3.6-27B-mmproj-BF16.gguf | ~888 MB (931,146,048 bytes) | Vision projector, passed through unmodified from llmfan46's repo |

The vision projector is required for multimodal (image/video) use.

It is identical to the file distributed by llmfan46 and is included here

for single-repo convenience only.

Provenance

1. Base architecture: Qwen/Qwen3.6-27B — Qwen Team (Apache-2.0)
2. Heretic variant: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF — llmfan46.

The BF16 GGUF from that repository was used as the direct quantization source.

llmfan46 applied Heretic v1.2.0 with the Magnitude-Preserving Orthogonal

Ablation (MPOA) method, targeting attn.o_proj, attn.out_proj, and

mlp.down_proj. Their reported result: 0.0021 KL divergence from base,

6/100 refusals vs 92/100 on the original model, MMLU accuracy 85.61%

vs 86.65% on the original.

1. Quantization: Cerebellum v4 recipe transferred verbatim from the stock

deucebucket/Qwen3.6-27B-Cerebellum-v4-GGUF

build — same 181-entry tensor-type override file, same coder imatrix

(ncall=8). The heretic build is verified to land within 4,400 bytes of the

stock v4 file (12,866,587,680 vs 12,866,583,200 bytes) — byte-class

identical.

Benchmarks

Benchmarks run on these GGUF files directly using llama.cpp on RTX 3090.

All numbers are audited; every failed answer was manually verified as a genuine

model error — audit reports are in benchmark_results/AUDIT_*.md.

Full per-question detail (summary JSON, samples JSONL, EvalPlus eval JSON,

adversarial audit reports) is in benchmark_results/ in this repository.

Heretic Cerebellum v1 (12.87 GB) vs baselines

| Benchmark | Heretic Cerebellum v1 (12.87 GB) | Stock Cerebellum v4 (12.87 GB) | Notes |

|-----------|:---:|:---:|---|

| Wiki PPL (ctx 2048, 32 chunks) | 6.925 ± 0.095 | 6.681 ± 0.091 | RTX 3090, identical invocation |

| ARC-Challenge | 96.93% (1172 q) | 96.76% | 25-shot |

| HellaSwag | 90.14% (10042 q) | 92.29% | 10-shot, stock re-measured same-night same-harness |

| MMLU-Redux | 76.21% (2400 q) | 76.58% | 5-shot |

| HumanEval base (chat, no-think) | 89.63% (164 problems) | 90.85% | pass@1, evalplus patched harness |

| HumanEval+ (chat, no-think) | 84.76% | 87.20% | pass@1, evalplus patched harness |

| Vision smoke | 100% (24/24) | — | basic image description |

| RealWorldQA | 78.0% (n=50, ±2%) | — | single-question granularity ±2% per question |

Stock Cerebellum v4 is the same tensor allocation applied to the non-heretic base.

HumanEval methodology note

The May-published "81.10%" for stock v4 was measured with thinking mode enabled

and is not comparable to the numbers above. The publishable comparison is

the same-night, same-harness, no-thinking chat pair on the patched harness

(zero fabricated completions, verified): heretic 89.63 base / 84.76 plus

vs stock v4 90.85 base / 87.20 plus.

For transparency, the raw-completions endpoint pair is also reported:

heretic 65.24% base / 60.98% plus vs stock v4 62.20% base / 56.71% plus.

Both raw-completions numbers were affected by a now-fixed harness timeout bug

(BE-15); the raw-completions endpoint under-serves this model family. Full

details are in benchmark_results/.

Heretic Abliteration Details (from llmfan46)

The following parameters are as reported in llmfan46's model card and are

reproduced here for downstream reference.

| Parameter | Value |

|-----------|-------|

| direction_index | 30.38 |

| attn.out_proj.max_weight | 1.58 |

| attn.out_proj.max_weight_position | 38.93 |

| attn.out_proj.min_weight | 1.51 |

| attn.out_proj.min_weight_distance | 32.78 |

| mlp.down_proj.max_weight | 1.80 |

| mlp.down_proj.max_weight_position | 41.28 |

| mlp.down_proj.min_weight | 0.54 |

| mlp.down_proj.min_weight_distance | 43.66 |

| attn.o_proj.max_weight | 1.99 |

| attn.o_proj.max_weight_position | 48.06 |

| attn.o_proj.min_weight | 1.75 |

| attn.o_proj.min_weight_distance | 39.00 |

Targeted components: attn.o_proj, attn.out_proj, mlp.down_proj.

Tool: Heretic v1.2.0,

method: Magnitude-Preserving Orthogonal Ablation (MPOA)

(reference).

Cerebellum v4 Tensor Allocation

Same allocation as the stock build. Listed here for reference.

181 sacred promotions over a Q2_K base: 70 Q5_K / 41 Q6_K / 7 Q8_0 /

22 Q4_K / 19 Q3_K / 22 Q2_K (explicit).

| Group | Precision | Rationale |

|-------|-----------|-----------|

| SSM state parameters | F32 | Hard-fail below 4-bit — NaN with no gradual degradation |

| SSM in_proj_a/b, A_log, dt_bias, conv1d, in_proj_z | ≥ Q4 minimum | 4-bit floor enforced per hybrid SSM ablation results |

| Most-sensitive attention tensors | Q5_K / Q6_K / Q8_0 | Sacred-pinned per per-tensor PPL ablation |

| Norm tensors | F32 | Protected; standard practice |

| Bulk ffn / remaining attention | Q2_K | Base precision, imatrix-guided |

Protected: all norms (F32), SSM recurrent state (F32).

Perplexity Note

Wiki PPL for the Heretic build (6.925 ± 0.095) is 0.244 higher than the stock

Cerebellum v4 (6.681 ± 0.091). The error bars do not overlap, indicating a

measurable distributional shift from the abliteration step. The delta is

consistent with what llmfan46 reported (MMLU drop of ~1 point) and is not

attributable to quantization differences — the two builds use identical

tensor allocations on a byte-class-identical file.

Runtime — Casual Deployment

llama-server \
  --model Qwen3.6-27B-Heretic-Cerebellum-v1-Q2_K_Mixed.gguf \
  --mmproj Qwen3.6-27B-mmproj-BF16.gguf \
  --n-gpu-layers 99 \
  --ctx-size 8192 \
  --jinja

--jinja is required for Qwen3.6. The enable_thinking chat-template flag

only takes effect when the Jinja template path is active; without it, the

model defaults to thinking mode on every request.

Non-thinking requests require an explicit flag at the API level:

{"chat_template_kwargs": {"enable_thinking": false}}

Qwen3.6 does not support the /think and /nothink soft-switch tokens

used by Qwen3.5. Thinking mode is on by default.

Recommended Sampling Parameters

From the official Qwen3.6-27B documentation.

| Mode | temperature | top_p | top_k | min_p | presence_penalty | repetition_penalty |

|------|-------------|-------|-------|-------|------------------|--------------------|

| Thinking — general | 1.0 | 0.95 | 20 | 0.0 | 1.5 | 1.0 |

| Thinking — precise coding (WebDev) | 0.6 | 0.95 | 20 | 0.0 | 0.0 | 1.0 |

| Non-thinking (instruct) | 0.7 | 0.80 | 20 | 0.0 | 1.5 | 1.0 |

presence_penalty can be adjusted between 0 and 2 to reduce repetition loops;

higher values may occasionally cause language mixing.

Reproduction

Standard Cerebellum recipe. The tensor-type override file and ablation logs

from the stock v4 build apply directly.

# 1. imatrix (constant ~300 MB RAM)
python -m osmosis.imatrix_stream \
    --model Qwen3.6-27B-uncensored-heretic-v2-BF16.gguf \
    --output imatrix.dat

# 2. quantize with stock llama-quantize
llama-quantize \
    --imatrix imatrix.dat \
    --tensor-type-file cerebellum_v4_overrides.txt \
    Qwen3.6-27B-uncensored-heretic-v2-BF16.gguf \
    Qwen3.6-27B-Heretic-Cerebellum-v1-Q2_K_Mixed.gguf \
    Q2_K

The imatrix used for this build was generated from the coder corpus

(ncall=8; same corpus as the stock Cerebellum v4 build).

The 181-entry tensor override file (cerebellum_v4_overrides.txt) is included

in this repository alongside the ablation logs.

Benchmark Artifacts

Summary JSONs, per-question JSONL samples, EvalPlus eval JSON files, and

adversarial audit reports (AUDIT_*.md) are in benchmark_results/ in this

repository per project policy.

Credits

• Base model: Qwen/Qwen3.6-27B — Qwen Team
• Heretic variant and BF16 source: llmfan46/Qwen3.6-27B-uncensored-heretic-v2-GGUF — llmfan46
• Abliteration tool: Heretic v1.2.0 by p-e-w
• GGUF runtime: llama.cpp
• Quantization method and workflow: Cerebellum — deucebucket