---

license: apache-2.0

library_name: gguf

pipeline_tag: text-generation

language:

• en
• zh

tags:

• qwen3.8
• heretic
• ara
• abliteration
• slerp
• merge
• mtp
• gguf

---

Qwen3.8-Smaug-Effithink-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF

Heretic-ARA (directional) abliterated variant of

Dalvlad/Qwen3.8-Smaug-Effithink-SLERP-27B,

quantized to IQ4_XS with the MTP block included — 14.36 GiB, sized

for 16 GB VRAM (same recipe family as Bucoid's Qwen3.8-27B Heretic-Ara

IQ4_XS-MTP GGUF on Hugging Face).

• Qwen3.8-Smaug-Effithink-Heretic-Ara-IQ4_XS-MTP.gguf — 14.36 GiB
• mmproj-Qwen3.8-Smaug-Effithink-SLERP-27B-Q8_0.gguf — vision projector

(Merkyor's official mmproj; the vision tower was untouched by ablation)

Method (honest version)

Ran heretic-llm 1.4.0 (p-e-w/heretic) on CPU — no GPU was available,

so instead of heretic's full Optuna search we ran a reduced budget:

6 parametrized ablation configs (per-layer + fractional-global refusal

directions, weight kernels across o_proj/down_proj) + an arbitrary-rank

(ARA-style) second pass with directions recomputed on the once-ablated

model. Scoring = heretic's own first-token KL divergence vs the base model +

refusal count on an AdvBench subset (n=10).

Results:

| model | refusals (AdvBench-10) | first-token KL |

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

| base merge (Qwen3.8-Smaug-Effithink-SLERP-27B) | 7/10 | 0 |

| Heretic-Ara (this model) | 4/10 | 0.0085 |

| round-2 stronger kernels (mw 1.2–1.5, flat) | 4–5/10 | 0.0020–0.0069 |

| round-2 ARA rank-2 pass | 6/10 (rejected) | 0.0041 |

Capability spot-checks after ablation: 17×24 − 5 = 403 ✓, Paris ✓,

no degradation observed in short coding answers.

Honest limitations — read before downloading

• This is a partial abliteration. Mild/professional-harm prompts

(lock picking, shoplifting) complied even before ablation (the

EfficientThink lineage is tagged "uncensored"), and the ARA pass flipped

roughly half of the remaining refusals. **Hard-safety categories —

weapons synthesis, drug manufacture — still refuse** ("I can't provide

step-by-step instructions for making a bomb", etc.).

• Refusal reductions plateaued at stronger kernel weights (1.2–1.5): the

refusal behavior of this RLOO-post-trained stack is not a

single-direction feature that rank-1/rank-2 ablation removes on a CPU

budget. A full uncensored result needs heretic's complete search on a

GPU (KL budget to ~0.03 was barely touched: we used 0.0085).

• KL 0.0085 means the model is nearly identical to its base — expect the

same intelligence, the same quirks, and about half the refusals gone.

Usage

llama-cli -m Qwen3.8-Smaug-Effithink-Heretic-Ara-IQ4_XS-MTP.gguf -c 262144 -t 16 -cnv
# vision: pair with the mmproj file above (same as the base merge)
llama-mtmd-cli -m Qwen3.8-Smaug-Effithink-Heretic-Ara-IQ4_XS-MTP.gguf \
  -mm mmproj-Qwen3.8-Smaug-Effithink-SLERP-27B-Q8_0.gguf --image <img> -p "..."

Credits

• p-e-w/heretic (heretic-llm, AGPL-3.0)

— the ablation machinery; "Heretic-Ara" follows Bucoid's

Arbitrary-Rank-Ablation naming.

• Bucoid/Qwen3.8-27B-Heretic-Ara-16GB-VRAM-IQ4-XS-MTP-GGUF

— the recipe this repo mirrors.

• Qwen/Qwen3.8-27B,

abacusai/Smaug-Mini,

Merkyor/Qwen3.8-27B-Coder390-EfficientThink-...

— the underlying models (Apache-2.0).

Showcase: hard-trace reasoning (verified outcome)

Challenged with a 7-trap CPython 3.13 semantics puzzle — late-binding

closures, memoized recursion with a cache side-effect, a generator with

yield in finally and a continue (finally re-runs → double

increment), ValueError propagation through list(), the implicit

del e after except ... as e, itertools.accumulate,

dict.get/setdefault — with an 8192-token thinking budget, graded

against real CPython output:

> 7/7 output lines exact, every mechanism explained correctly.

It used ~4k tokens of the thinking budget (didn't need the ceiling) and

self-verified each line before committing — e.g. it enumerated all 9

memo-cache entries and derived the finally-on-continue double

increment (…31, then 32) rather than pattern-matching. At IQ4_XS on 16

CPU threads: 4.2 tok/s generation, 27.5 t/s prompt processing.