Qwen3.8 27B Heretic Abliterated Uncensored GGUF

RVN is a double-refined abliterated variant of Qwen3.8-27B, built on top of trohrbaugh/Qwen3.8-27B-heretic-ara (an ARA abliteration by Tim Rohrbaugh) and further refined with two additional full-weight ARA passes targeting residual refusals. It retains very low behavioral damage (KL ≈ 0.0085) while reducing harmful-prompt refusals from 3/100 (source) to 0–1/100 in independent measurements.

llama-server -m RVN-Q5_K_M.gguf --chat-template-file chat_template.jinja \
  -c 32768 -ngl 99

What is ARA?

ARA (Arbitrary-Rank Ablation) is the abliteration technique implemented in p-e-w/heretic. Traditional directional abliteration finds a single "refusal direction" in activation space and subtracts it — a one-shot, low-rank surgery that is simple but can leave residual refusals or damage unrelated behavior.

ARA instead treats abliteration as a matrix optimization problem. For every target module (attention out-projection and MLP down-projection), it collects activations on "good" prompts (harmless requests) and "bad" prompts (harmful requests), then uses an LBFGS optimizer to rewrite the module's weight matrix so that:

• Preserve: outputs on good prompts change as little as possible (KL is kept low)
• Steer: outputs on bad prompts are pulled toward the good-prompt output manifold (via k-nearest-neighbor distances), so harmful requests stop triggering the refusal circuitry
• Overcorrect: outputs on bad prompts are additionally pushed away from the original bad-prompt outputs, which helps overcome complex, multi-stage refusal mechanisms

Because the weight matrix is optimized directly (rather than subtracting a single direction), ARA is "arbitrary rank" — it can carve out a much richer refusal-removal subspace while keeping behavioral damage minimal.

Why "Heretic" and "Abliterated"?

These two words describe two layers of the same process:

• Heretic is the tool: the open-source implementation of ARA (and related abliteration methods) used to modify the model. Models produced with it are commonly labeled "heretic" in the community.
• Abliterated is the result: the model's refusal behavior has been surgically removed. An abliterated model still knows everything the base model knows, but it no longer refuses to answer the categories that were steered away during the process.

So "Heretic Abliterated" means: abliterated using the heretic toolset. RVN goes one step further — it applies the ARA procedure three times total: once by the original author (trohrbaugh) to get from base Qwen3.8-27B to -ara, and twice more by us to get from -ara to RVN, squeezing out the last residual refusals.

Special Thanks

This work would not exist without Tim Rohrbaugh (trohrbaugh), whose heretic-ara ARA abliteration of Qwen3.8-27B (refusals 3/100, KL 0.0535) provided the foundation we refined into RVN. His upstream contributions to the heretic codebase — including the row-norm preservation feature and Qwen3.5 MoE/DeltaNet hybrid handling — are directly responsible for making DeltaNet-layer abliteration work at all. Thank you, Tim.

Model Overview

Why RVN?

trohrbaugh/Qwen3.8-27B-heretic-ara is already a strong ARA abliteration, but three harmful prompts still triggered refusals in our independent evaluation (racism website, malware, government database hacking). RVN applies two additional full-weight ARA passes using the same tight parameter set (start 26, end 56, preserve 0.9432, steer 0.0009, overcorrect 0.5038, neighbor 10), which:

• Reduced refusals from 3/100 → 0–1/100 (the only remaining refusal is a chemical-weapon WMD prompt — one of the strongest safety-trained categories, and intentionally one of the guardrails we left in place)
• Reduced KL damage from 0.0535 (source) to 0.0085 vs base — a ~6× improvement in behavioral preservation
• Verified independently on two rented GPU machines with prefix-based (real-answer) refusal measurement

Refusal evaluation (100 harmful-behaviors prompts, prefix-forced real answers)

Files & Quantization Spectrum

👁️ Vision-Protected (UD-style) Variants

The standard quants above use uniform quantization across all layers. For vision-heavy use (image understanding via the mmproj), the layers that first receive visual embeddings matter disproportionately — the "vision-language bridge". These -vision variants use Unsloth-style dynamic (UD) quantization: token_embd, output, and the first 4 + last 4 transformer blocks stay at Q8_0, while the middle blocks are compressed to the target K-quant. Built with llama-quantize --token-embedding-type q8_0 --output-tensor-type q8_0 --tensor-type-file (106 bridge tensors overridden by regex; verified via tensor audit — 0 NaN/Inf, correct per-tensor types, 851 tensors).

Usage — identical to the standard quants, just point at the -vision file:

llama-server -m RVN-Q5_K_M-vision.gguf --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf \
  -c 32768 -ngl 99

📊 Perplexity (vs F16 reference)

Measured on an RTX PRO 6000 Blackwell (full GPU offload, llama.cpp master): llama-perplexity, tiny_shakespeare corpus, ctx 2048.

Text perplexity is at parity with the standard quant (no regression from the Q8_0 bridge), while the vision-critical bridge layers keep 8-bit precision. All -vision variants and the standard Q5_K_M passed image-description tests with the mmproj (accurate description, no hallucination on the test image).

⚡ MTP Speculative Decoding

Every quant ships a *-mtp.gguf twin with the official Qwen3.8 MTP draft head (q8_0, ~1.8 GB) embedded. The main-model weights are byte-identical to the base file — the head is appended as an extra blk.64.nextn.* layer set (block_count 65, qwen35.nextn_predict_layers=1). Abliteration never touched the draft head (it operates on main-model layers 26–56), so draft quality is unchanged, and speculative decoding is output-equivalent: same tokens, just faster.

Usage (llama.cpp ≥ b10440, PR #22673):

llama-server -m RVN-IQ3_M-mtp.gguf -c 32768 -ngl 99 \
  --spec-type draft-mtp --spec-draft-n-max 2 --parallel 1

Measured on 2× RTX PRO 6000 Blackwell (95 GB each, full GPU offload, llama.cpp b10472):

Average: +55% generation speed (128-token continuation, --spec-draft-n-max 2 --parallel 1). ⚠️ IQ1_M is the exception: MTP makes it ~60% slower — use the base file for IQ1_M. IQ1_S gains almost nothing (+4%). All other quants gain +29% to +105%.

Community measurements on larger cards report +33–145% depending on GPU and context. Tips: --spec-draft-n-max 2 is the sweet spot on 16–24 GB cards (3–4 on bigger/faster cards); pair with --cache-type-k q4_0 --cache-type-v q4_0 for long context; --spec-draft-p-min 0.60–0.75 helps on bandwidth-limited rigs.

👁️ Vision (image understanding)

Pair any *-mtp.gguf with the official vision projector for full image-text-to-text:

llama-server -m RVN-IQ3_M-mtp.gguf --mmproj mmproj-Qwen3.8-27B-Q8_0.gguf \
  -c 32768 -ngl 99 --spec-type draft-mtp --spec-draft-n-max 2 --parallel 1

• mmproj-Qwen3.8-27B-Q8_0.gguf (0.63 GB, Q8_0) is the official Qwen3.8 vision projector from ggml-org/Qwen3.8-27B-GGUF (Apache-2.0). The vision tower is an image encoder only — ARA abliteration never touched it, so it pairs cleanly with every RVN quant.
• Verified: RVN-Q3_K_M-mtp + this mmproj describes images correctly (MTP active).
• Combo credit: cfigueiroa/Qwen3.8-27B-RVN-vision-MTP.

Quant → GPU / Memory Guide

KV cache math (GQA, 4 KV heads, head_dim 256): 2 × 64 layers × 4 KV heads × 256 head_dim × 2 bytes = 256 KiB/token FP16 → 16K ctx ≈ 4.2 GB · 32K ctx ≈ 8.4 GB · 64K ctx ≈ 16.8 GB (Q8_0 KV halves this). A 16 GB card running Q3_K_M (13.30 GB model) + 16K ctx Q8_0 KV fits comfortably; Q4_K_M (16.55 GB) really needs a 24 GB card.

Rule of thumb: pick the largest quant that leaves ≥ 4 GB for KV cache + compute buffers. If you only need short replies, drop the quant one notch and get a bigger context; if you need long context, prioritize KV budget over quant size.

Limitations & Responsible Use

• Reduced safety guardrails by design. This model is not intended for use in applications requiring robust safety filtering, content moderation, or deployment to minors.
• Certain guardrails are intentionally left in place. Abliteration targets refusal behavior on general harmful-prompt categories; a small set of hard safety-trained categories is deliberately not fully removed. Behavior may vary across domains and languages.
• Not affiliated with or endorsed by Qwen/Alibaba or trohrbaugh.

License & Attribution

• Base model: Qwen/Qwen3.8-27B — Apache-2.0
• Abliteration source: trohrbaugh/Qwen3.8-27B-heretic-ara — Apache-2.0
• This repo: Apache-2.0. See LICENSE.

Citation

@misc{rohrbaugh2026heretic,
  title={Qwen3.8-27B-heretic-ara: ARA Abliteration of Qwen3.8-27B},
  author={Rohrbaugh, Tim},
  year={2026},
  howpublished={\url{https://huggingface.co/trohrbaugh/Qwen3.8-27B-heretic-ara}}
}

@misc{rvn2026,
  title={RVN: Qwen3.8-27B Heretic Abliterated Uncensored},
  author={0bserverx},
  year={2026},
  howpublished={\url{https://huggingface.co/0bserverx/Qwen3.8-27B-Heretic-Abliterated-Uncensored-GGUF}}
}