---

license: apache-2.0

library_name: gguf

pipeline_tag: text-generation

base_model:

• llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved

tags:

• gguf
• llama.cpp
• qwen3
• abliterated
• uncensored
• heretic
• unsloth
• imatrix
• dynamic-quant
• mtp

---

Qwen3.8-27B-Heretic-Dynamic-GGUF

Dynamic (Unsloth-recipe) GGUF quants of the Heretic-abliterated Qwen3.8-27B.

Currently available: UD-Q4_K_XL, UD-Q5_K_M, UD-Q6_K_XL, UD-Q8_K_XL. More

sizes may be added if there is interest — each is built the same way described below

(Unsloth per-tensor layout + imatrix on the Heretic BF16).

These quants combine two things that were previously only available separately:

• the Heretic MPOA abliteration from llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved (uncensored, MTP-preserved), and
• Unsloth's Dynamic v3.0 per-tensor quantization layout + importance matrix, which the source repo's own quants (plain static K-quants) do not use.

The result is measurably closer to the full-precision model than a static K-quant of the

same weights — verified by KL divergence at Q5 (below) — at a slightly smaller file size.

What this is, precisely

• Weights: llmfan46's Heretic BF16 (Heretic v2.0.0.dev0, a variant of Magnitude-Preserving

Orthogonal Ablation). All 15 MTP tensors preserved; vision projector untouched.

• Quantization: llama-quantize --allow-requantize with Unsloth's published

imatrix_unsloth.gguf and an exhaustive per-tensor --tensor-type-file generated from

the corresponding Unsloth GGUF of the original model. Each size's tensor-type layout is

verified identical to its Unsloth counterpart (866/866 tensors, 0 type mismatches).

Tensor-type distributions (F32 norms etc. excluded):

| Size | Distribution |

| :--- | :--- |

| UD-Q4_K_XL | 191×Q5_K, 110×Q8_0, 70×IQ4_XS, 69×Q4_K, 56×Q6_K, 6×IQ4_NL, 3×Q3_K, 1×IQ3_S |

| UD-Q5_K_M | 189×Q5_K, 160×Q6_K, 124×Q8_0, 19×IQ4_XS, 12×Q4_K, 2×IQ4_NL |

| UD-Q6_K_XL | 310×Q8_0, 168×Q6_K, 27×Q5_K, 1×Q4_K |

| UD-Q8_K_XL | 453×Q8_0, 53×BF16 |

Why dynamic over the static Q5 — measured

KL divergence vs a non-abliterated Unsloth Q8_0 reference (wikitext-2 test, -c 8192,

f16 KV; self-control floor: mean KLD 0.000000, top-1 99.994%). Because both Q5s carry the

identical abliteration, the gap between them isolates the quantization method:

| Q5 of the Heretic model | Mean KLD | Median KLD | 99% KLD | Same top-1 | File size |

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

| This repo (Unsloth dynamic) | 0.070421 | 0.008692 | 0.898 | 92.327 % | 18.4 GB |

| Static Q5_K_M (source repo) | 0.076442 | 0.010440 | 0.985 | 91.886 % | 18.7 GB |

Dynamic is lower divergence on every axis — mean, median, tail, top-1 agreement — in a

smaller file. (For reference, Unsloth's dynamic Q5 of the original non-abliterated model

scores 0.021978 mean KLD / 96.933 % top-1 against the same base; the remainder of this

model's divergence is the intended abliteration, not quantization loss.)

Files

| File | Size | Notes |

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

| Qwen3.8-27B-Heretic-Dynamic-UD-Q4_K_XL.gguf | 16.4 GiB | Smallest — fits a 24 GB GPU with lots of context. |

| Qwen3.8-27B-Heretic-Dynamic-UD-Q5_K_M.gguf | 18.4 GiB | The KLD-measured quant below. |

| Qwen3.8-27B-Heretic-Dynamic-UD-Q6_K_XL.gguf | 23.6 GiB | Best fit for a 32 GB GPU. |

| Qwen3.8-27B-Heretic-Dynamic-UD-Q8_K_XL.gguf | 29.3 GiB | Near-lossless (Q8_0 + BF16 tensors); needs >32 GB VRAM or partial offload. |

| mmproj-BF16.gguf | 0.9 GiB | Vision projector — required only for image/video input. Unchanged from the base model (abliteration and quantization do not touch it). |

All main GGUFs contain the 15 MTP tensors.

Qwen3.8-27B is a native vision-language model; the vision encoder lives in the separate

mmproj-BF16.gguf. It is optional — text-only use needs just the main GGUF.

Usage (llama.cpp)

# text only
llama-cli -m Qwen3.8-27B-Heretic-Dynamic-UD-Q5_K_M.gguf -c 8192

# with vision (image/video input)
llama-mtmd-cli -m Qwen3.8-27B-Heretic-Dynamic-UD-Q5_K_M.gguf --mmproj mmproj-BF16.gguf -c 8192

Sampling (from Qwen3.8's own recommendation, thinking mode):

temperature 1.0, top_p 0.95, top_k 20, min_p 0.0.

Reproduce

# per-tensor layout copied from Unsloth's GGUF of the same size, imatrix from Unsloth's repo
# (fallback base type: Q4_K_M / Q5_K_M / Q6_K / Q8_0 to match the size being built)
# 1. make the per-size --tensor-type-file from the combined map (here: UD-Q5_K_M)
awk -F'\t' 'NR==1 { for (i=2; i<=NF; i++) if ($i=="UD-Q5_K_M") c=i; next }
            { n=$1; gsub(/\./,"\\\\.",n); print "^" n "$=" $c }' \
  tensor_types.tsv > tensor_types_UD-Q5_K_M.txt

# 2. quantize
llama-quantize --allow-requantize \
  --imatrix imatrix_unsloth.gguf \
  --tensor-type-file tensor_types_UD-Q5_K_M.txt \
  Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved-BF16.gguf \
  Qwen3.8-27B-Heretic-Dynamic-UD-Q5_K_M.gguf Q5_K_M

tensor_types.tsv (in this repo) holds the per-tensor type for every non-F32 tensor in

every size, read from the GGUF headers of the corresponding Unsloth files — one row per

tensor, one column per size. The generated regexes must be anchored (^…$) —

llama-quantize matches them unanchored, and a bare output\.weight line would silently

clobber every attn_output.weight.

Note: the imatrix was calibrated on the original (non-abliterated) model. Given how close

the abliterated model is to the original, it remains a valid importance proxy — the same

assumption Unsloth's and other abliterated dynamic quants rely on.

Credits

• Original model: Qwen/Qwen3.8-27B (Apache-2.0)
• Abliteration + BF16 source: llmfan46/Qwen3.8-27B-Ultra-Uncensored-Heretic-Native-MTP-Preserved, made with Heretic
• Dynamic quant recipe + imatrix: unsloth/Qwen3.8-27B-GGUF

Usage warnings

This is an uncensored model: its safety filtering has been substantially removed and it

can produce sensitive, controversial, or otherwise inappropriate content. Review outputs,

use it in controlled settings, and ensure your use complies with applicable law and the

Apache-2.0 license of the base model. Responsibility for generated content rests with the

user.