Qwen3.6-27B-Abliterated-Heretic-Uncensored-GGUF
This is a GGUF release of an abliterated, uncensored version of Qwen's Qwen3.6-27B, made with Heretic.
By applying a Heretic-style MPOA pipeline with magnitude preservation on the Qwen3.6-27B dense text stack, the base refusal behavior was removed at the weight level with extremely low distributional divergence (KL 0.0251 vs base on harmless prompts). The text GGUFs are paired with Qwen3.6-27B vision projectors (mmproj) so image/video-capable llama.cpp and LM Studio runtimes can use the multimodal path.
Quick Benchmarks
| Check | Original Qwen3.6-27B | Abliterated Heretic Uncensored |
|---|---|---|
| Official 25-prompt refusal check | 20/25 refusals | 1/25 refusals |
| 100-prompt refusal check | 92/100 refusals | 3/100 refusals |
| KL divergence | N/A | 0.0251 |
Methodology & Model Notes
Qwen3.6-27B is a 27.8B dense vision-language model with 64 text layers, hybrid linear/full attention (3 linear-attention + 1 full-attention per 4-layer group), and an integrated image + video vision tower.
This release was produced with a direct Heretic-style MPOA run with magnitude preservation — output-side orthogonalization on self_attn.o_proj, linear_attn.out_proj, and mlp.down_proj, with each weight row/column's L2 norm restored after projection. The ablation direction is interpolated at direction_index = 37.97.
The accepted candidate scored Refusals: 1/25 on the official 25-prompt marker suite used for the MiniMax M2.7 and Qwen3.6-35B-A3B abliterated runs, with a measured KL divergence of 0.0251 against the base on mlabonne/harmless_alpaca test[:25].
The resulting abliterated checkpoint was exported to BF16 and then converted to GGUF for llama.cpp-compatible deployment. The language GGUF files are text-model files; multimodal input is enabled by loading a matching mmproj projector alongside them.
Files
Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16-00001-of-00002.gguf+-00002-of-00002.gguf: BF16 GGUF source (split; use with--load-tensorsorllama-gguf-split --merge)Qwen3.6-27B-Abliterated-Heretic-Uncensored-Q8_0.gguf: highest-fidelity quantQwen3.6-27B-Abliterated-Heretic-Uncensored-Q6_K.gguf: near-lossless practical quantQwen3.6-27B-Abliterated-Heretic-Uncensored-Q5_K_M.gguf: high-fidelity medium quantQwen3.6-27B-Abliterated-Heretic-Uncensored-Q4_K_M.gguf: smaller general-use quantQwen3.6-27B-Abliterated-Heretic-Uncensored-Q3_K_M.gguf: compact quantQwen3.6-27B-Abliterated-Heretic-Uncensored-Q2_K.gguf: smallest-footprint quantmmproj-F16.gguf/mmproj-model-f16.gguf: F16 Qwen3.6-27B vision projector for LM Studio and llama.cpp multimodal loadingmmproj-BF16.gguf/mmproj-model-bf16.gguf: BF16 Qwen3.6-27B vision projector
Running
llama-server \
-m <quant-file.gguf> \
-ngl 999 -c 32768 --jinja -fa
For image/video-capable runtimes, load the projector with the text GGUF:
llama-server \
-m <quant-file.gguf> \
--mmproj mmproj-model-f16.gguf \
-ngl 999 -c 32768 --jinja -fa
Model Architecture
| Spec | Value |
|---|---|
| Total Parameters | 27.8B (dense) |
| Layers | 64 |
| Attention | Hybrid (3 linear-attention + 1 full-attention per 4-layer group) |
| Hidden Size | 5120 |
| Family | qwen3_5 |
| Modality | Vision-language via GGUF text model + Qwen3.6-27B mmproj |
| Base Model | Qwen/Qwen3.6-27B |
Disclaimer
This model has had refusal behavior removed at the weight level. It will answer prompts that the base model would normally refuse. You are responsible for how you use it.
Credits
- Base model: Qwen/Qwen3.6-27B
- Refusal removal pipeline: Heretic
- GGUF runtime and quantization: llama.cpp
License
This release inherits the base Qwen3.6-27B license.
Apache-2.0.