BF16 abliterated Qwen release
According to the model card, this is a BF16 release of an abliterated, uncensored version of Qwen3.6-27B with vision, derived from Qwen/Qwen3.6-27B with the Heretic pipeline.
Open Source Model Profile · Youssofal
Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16
Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16 is a 27.36B-parameter Qwen vision-language release from Youssofal. According to the model card, it is a BF16 abliterated version with refusal behavior attenuated.
- Publisher
- Youssofal
- Task
- image-text-to-text
- Model type
- qwen3_5
- License
- apache-2.0
- Library
- transformers
- Publication status
- Accepted · not indexed
Model overview
Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16 is published by Youssofal as a qwen3_5 image-text-to-text model. The captured configuration identifies Qwen3_5ForConditionalGeneration with model type qwen3_5, and Safetensors metadata reports 27,356,728,560 parameters. According to the model card, it is a BF16 abliterated, uncensored derivative of Qwen3.6-27B with vision.
Recorded capabilities
27.36B Qwen3_5 weights
Captured configuration identifies Qwen3_5ForConditionalGeneration with model type qwen3_5, and Safetensors metadata reports 27,356,728,560 parameters.
Two-stage MPOA ablation
According to the model card, a Heretic-style two-stage MPOA pipeline with magnitude preservation applied output-side and jailbreak-conditioned input-side ablation.
Vision-video architecture notes
According to the model card, the base is described as a 64-layer dense vision-language model with hybrid linear and full attention, 5120 hidden size, and image-plus-video support.
Use cases in the source record
- Vision-language experimentation with the publisher-documented image and video input path, using AutoProcessor loading with trust_remote_code.
- Research into refusal-attenuated behavior where the publisher notes the model answers prompts the base model would refuse, with user responsibility for use.
Limitations and unknowns
- No Ethen-measured evaluation results were extracted from this record.
- No context-window value was extracted.
- Provider state is historical snapshot data and should be refreshed before being presented as current.
- Ablation, benchmark, and capability claims are publisher model-card statements and were not independently verified by Ethen.
Source and provenance
Source: Youssofal/Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16
Captured: Unknown. Processed: 2026-09-07T19:35:53.096886+00:00.
Qwen3.6-27B-Abliterated-Heretic-Uncensored-BF16 This is a BF16 release of an abliterated, uncensored version of Qwen's Qwen3.6-27B with vision, made with Heretic. By applying a Heretic-style two-stage MPOA pipeline with magnitude preservation on the Qwen3.6-27B dense text stack — slot-grouped output-side ablation followed by jailbreak-conditioned input-side ablation, mirroring the methodology used for Qwen3.6-35B-A3B — the base refusal behavior was attenuated at the weight level with low distributional divergence (KL 0.0282 vs base on harmless prompts). The result keeps Qwen3.6-27B's full vision and video multimodal architecture and…
F001F002F003F004F005F006F007F010F011F012F013F014F015F017F018