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`](https://huggingface.co/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 variant and is superseded by the RVN files below.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. **ARA (Arbitrary-Rank Ablation)** is the abliteration technique implemented in [p-e-w/heretic](https://github.com/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. 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. This work would not exist without **Tim Rohrbaugh** (`trohrbaugh`), whose [`heretic-ara`](https://huggingface.co/trohrbaugh/Qwen3.8-27B-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. `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 **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. The sizes below are the live model-weight sizes at the verified repository head. They do **not** include runtime buffers or KV cache. 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. - **`-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. - **`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. 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. - 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. - 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. 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`](https://gist.github.com/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):