Qwen3.8-27B RVN 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.

Note on this repository's history. This repo previously hosted the original Qwen3.8-27B-Heretic-Q4_K_M.gguf (single-quant release from the earlier trohrbaugh/Qwen3.8-27B-heretic source). That file is kept as legacy for download-count continuity and backward compatibility — it is the older abliteration variant and is superseded by the RVN files below. Prefer the RVN quants for new deployments.

Not for all audiences. This model has reduced safety guardrails by design. It is intended for adult audiences (18+) doing research, creative writing, roleplay, and uncensored generation. Certain guardrails are intentionally left in place; use responsibly and in accordance with your local laws.

Compatibility status (2026-08-19): the recommended *-multilingual*.gguf family and all 53 legacy RVN GGUF paths embed the official Qwen3.8 chat template. Every multilingual artifact passed a real per-file OpenAI-compatible tool-call/thinking-control gate; the legacy paths were repaired in place so users do not need an external template workaround.

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)

Which file should I download?

Download one main-model GGUF, not the whole repository. Start with a plain -multilingual.gguf file unless you specifically need MTP or a vision-protected build. Model size is not the whole memory requirement: leave room for the KV cache, compute buffers, the operating system, and—when used—the vision projector.

Quick picks

The sizes below are the live model-weight sizes at the verified repository head. They do not include runtime buffers or KV cache.

The valid filename families

Suffixes cannot be combined arbitrarily. These are the families that actually exist:

RVN-{QUANT}.gguf
RVN-{QUANT}-mtp.gguf
RVN-{Q5_K_M|Q4_K_M|Q3_K_M}-vision.gguf

RVN-{QUANT}-multilingual.gguf
RVN-{QUANT}-multilingual-mtp.gguf
RVN-{Q5_K_M|Q4_K_M|Q3_K_M}-multilingual-vision.gguf

RVN-F16.gguf / RVN-F16-mtp.gguf
RVN-BF16.gguf / RVN-BF16-mtp.gguf

mtp-RVN.gguf  # standalone draft-head compatibility artifact; not a main model

• There is no combined vision-mtp file.
• There is no F16-multilingual or BF16-multilingual file.
• mmproj-Qwen3.8-27B-Q8_0.gguf is a separate vision projector, not another main model.
• mtp-RVN.gguf is a separate 1.69 GiB standalone MTP draft-head artifact retained for advanced/legacy workflows. It cannot answer prompts by itself and is not required by any embedded *-mtp.gguf twin. Most users should ignore it and download either one plain main model or its embedded-MTP twin.

What the names mean

• -multilingual — the recommended Qwen-tuned calibration family for new downloads. Its importance matrix covers Turkish, Russian, 20+ other languages, code, reasoning, and tool-use structures instead of concentrating the low-bit precision budget on narrow English-only text. Calibration matters most at low bit-rates; it does not teach Qwen new languages.
• -mtp — the same main-model tensor payload plus Qwen3.8's official 15-tensor MTP/NextN draft head for speculative decoding. Budget about 451 MB / 0.42 GiB of additional memory. Every current multilingual twin is exactly 451,320,768 bytes larger; current legacy twins add either 451,320,736 or 451,320,768 bytes because of a 32-byte historical metadata-layout difference. It is a speed feature, not a higher-quality quant.
• -vision — keeps token embeddings, output, and the first/last bridge blocks at higher precision. It is an optional bridge-preserving variant, not a requirement for images. Multilingual vision files passed the repository's API/template gates; their image-quality advantage has not been separately benchmarked, so no quality uplift is claimed here.
• No -multilingual in the name — the original RVN calibration family, retained under its established filenames for compatibility and download continuity. It includes plain, -mtp, and selected -vision files. These current files are usable and template-correct, but the multilingual family is recommended for new low-bit deployments.

Quant label legend

• Q…_K… — llama.cpp K-quant family. Within the same tier, _M generally retains more precision and is larger than _S; _L is larger again where available.
• IQ… — importance-aware low-bit quant format. It is especially useful when memory is tight, but the filename alone is not proof of how a historical artifact was calibrated; this repository documents provenance separately.
• L → M → S → XS → XXS — within a comparable family, generally moves toward smaller size and lower fidelity.
• NL — nonlinear IQ4 variant.
• Do not rank unlike families from the suffix alone. Use the quick picks and exact live-size table below.

Vision: what to download

For image input, download:

1. One compatible RVN main-model GGUF—a standard or -vision file; and
2. mmproj-Qwen3.8-27B-Q8_0.gguf separately (about 0.63 GB).

Any compatible RVN main model can use the projector. Choose a -vision build when the available Q5_K_M/Q4_K_M/Q3_K_M size fits and preserving extra precision in the vision-language bridge is important. Because there is no combined vision-mtp artifact, choose between the self-contained vision-protected build and an embedded-MTP build.

MTP: when to choose it

• Start with the plain -multilingual.gguf file for maximum runtime compatibility.
• Choose the matching -multilingual-mtp.gguf twin only when your recent llama.cpp build supports embedded MTP, the extra 0.42 GiB still leaves enough runtime/context memory, and you enable the documented MTP flags.
• Measured speedup varies by quant and hardware. It is not guaranteed.
• Do not use the MTP twin for IQ1_M: measured generation was about 60% slower. IQ1_S gained only about 4%. Prefer the plain files for those extreme quants unless your own benchmark shows otherwise.

Legacy-name and chat-template clarification

• All 53 current RVN-*.gguf paths and all 49 *-multilingual*.gguf artifacts embed the official Qwen3.8 chat template. No external chat-template launch flag is needed for these current files.
• Qwen3.8-27B-Heretic-Q4_K_M.gguf is a separate, older pre-RVN abliteration retained for historical download continuity. Its current live header also contains the official embedded template, but it is not the RVN model and was not part of the 53-file RVN runtime-repair campaign. New users should choose an RVN multilingual file instead.
• chat_template.jinja remains in the repository as a readable reference copy of the official template, not as a required workaround for current files.

Files & Quantization Spectrum

Recommended: Qwen Multilingual v5 calibrated family

These additive files were rebuilt from the template-correct RVN F16 reference using a pinned Qwen-tuned multilingual/code calibration corpus. The legacy filenames remain available for download continuity; choose the -multilingual family for new deployments.

Why the multilingual family matters: Qwen already learned its languages during pretraining; calibration does not teach them again. The importance matrix controls which activation/weight regions are preserved most carefully during low-bit quantization. A narrow English-only calibration set can spend that precision budget on English encyclopedia patterns, while this Qwen-tuned corpus also represents Turkish, Russian, 20+ other languages, code, reasoning traces, and tool-use structures. That makes -multilingual the recommended default for new downloads, especially at 4-bit and below.

Which file should I choose? Use the plain -multilingual.gguf file for maximum runtime compatibility. Choose its -multilingual-mtp.gguf twin only with a recent llama.cpp build and the documented MTP flags when you want speculative-decoding speed. Choose a -multilingual-vision.gguf variant when image understanding with the official mmproj is a priority. Q8_0 is retained in the same family for a consistent, fully verified provenance path even though calibration effects are expected to be smaller at 8-bit precision.

Q8_0 clarification: full tensor-schema and tensor-payload hashing proved that the legacy and multilingual Q8_0 base files have byte-identical tensors; their MTP twins are likewise tensor-identical. The multilingual Q8_0 names are retained for family/provenance consistency, not as a claim of extra Q8 quality. Their GGUF metadata records the pinned multilingual imatrix provenance, while -mtp still denotes the separate 15-tensor speculative-decoding head. Comparison receipt SHA-256: 210e59a07412df8f46b94f41790c0fb36540a2a6a09fb24933e6c1c3cb1ee863.

Reproducibility and verification:

• Calibration source: tristandruyen/9e207a95c7d75ddf37525d353e00659c at revision aba17fe897c00fae02a18d26068aa453dee09e50.
• Corpus SHA-256: 2a0118c633250aad448cfc2d103e10a731af6974dab1738ca19261fce3587b80; imatrix SHA-256: 5e73e1449451480fa48f463b781cfa2124daaf5c337a0c44f272e37b64fc1114.
• Imatrix: 55 completed chunks, 496 entries, context 2048; llama.cpp commit 645ca2834bc16e7eab112a91aeb282ebb913f935.
• Embedded official Qwen template SHA-256: c3cf9e34abf4f9e36c2d72165aa9c132d3e2a725b6c2586aaa3a8af9d7a81041.
• 49/49 artifacts passed live LFS SHA-256 and 64 MiB header verification after upload.
• Every artifact was loaded in llama.cpp and then served through llama-server; its OpenAI-compatible API returned HTTP 200, finish_reason=tool_calls, and add({"a":19,"b":23}). enable_thinking:false suppressed <think>, and a separate normal-chat control returned TEMPLATE_OK.
• MTP twins additionally load with 866 tensors, qwen35.block_count=65, and qwen35.nextn_predict_layers=1.

Pilot PPL deltas (multilingual - reference; lower is better):

Vision-protected multilingual variants retain Q8_0 embeddings/output and the first/last bridge layers while quantizing the middle blocks:

👁️ Vision-Protected (UD-style) Variants

These -vision variants apply a structural precision policy: token_embd, output, and the first 4 + last 4 transformer blocks stay at Q8_0, while the middle blocks use the target K-quant. A tensor audit verified 106 regex-selected overrides, 851 tensors, correct per-tensor types, and 0 NaN/Inf. This describes the artifact structure; it is not a claim of a separately benchmarked image-quality uplift.

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.

The table above reports text perplexity only. The Q8_0 overrides are structurally verified, but this repository does not publish a reproducible comparative image-quality receipt for these files; no image-quality uplift is claimed.

imatrix-based quants are produced from the same F16 with an activation importance matrix computed over wikitext-2-raw (original spectrum, 580 chunks) or tiny_shakespeare (2026-08-17 re-quant additions: IQ3_M fix + IQ2_S/IQ3_XXS/IQ3_XS/IQ3_S, 159 chunks, llama-imatrix, -ngl 99). The currently published IQ4_XS and IQ4_NL files are exceptions: an independent GGUF-header audit and the retained production script confirm they were made without --imatrix. Their earlier imatrix labels were incorrect; corrected wikitext-imatrix replacements (including MTP twins) are being rebuilt. The -vision files are quantized with K-quant defaults (no imatrix) — the bridge protection is structural (Q8_0 overrides), not imatrix-dependent.

⚡ MTP Speculative Decoding

Every quant ships a *-mtp.gguf twin with the official Qwen3.8 MTP draft head (Q8_0; budget about 451 MB / 0.42 GiB per embedded twin) 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).

The separate mtp-RVN.gguf file is a 1.69 GiB standalone draft-head compatibility artifact for advanced/legacy workflows. It is not a main model, cannot answer prompts by itself, and is not required by the embedded *-mtp.gguf twins. 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

The values below are model weights only. KV cache, compute buffers, the operating system, MTP, and the vision projector need additional memory. For a new deployment, prefer the -multilingual filename shown here.

Budget about 451 MB / 0.42 GiB for an embedded MTP head. Current multilingual twins add exactly 451,320,768 bytes; current legacy twins add either 451,320,736 or 451,320,768 bytes because of a 32-byte historical metadata-layout difference. The separate vision projector is about 0.63 GB. Budget both on top of the main-model 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}}
}

Research use and responsibility

This repository is intended for legitimate research and controlled evaluation, including interpretability, alignment and refusal-behavior analysis, red-team testing, and robustness work. It is not a ready-made production safety layer. If you deploy the model or expose it to other users, you are responsible for adding suitable access controls, moderation, monitoring, and abuse prevention.

Use of these files is subject to the Apache License 2.0 and all applicable laws. You are responsible for how you operate the model and for outputs produced in your environment. To the extent permitted by law, the repository maintainers and upstream authors accept no liability for misuse or resulting harm. Generated outputs are not statements or endorsements by the maintainers, upstream creators, or their organizations.