Qwen3.8-27B-heretic

Model overview

darkc0de publishes Qwen3.8-27B-heretic as a transformers image-text-to-text model. Captured config identifies Qwen3_5ForConditionalGeneration with a qwen3_5 model type, and Safetensors metadata reports 27,356,728,560 parameters. Card data records apache-2.0. Hub tags include heretic, uncensored, decensored, abliterated, reproducible, and conversational. The model card says the conversion is reproducible and points to a reproduce README.

Recorded capabilities

Use cases in the source record

• Image-text-to-text generation on the captured pipeline tag, including conversational serving listed in the snapshot.
• Publisher-described vision-language use such as image and video understanding, attributed only to the captured card.

Limitations and unknowns

• No independently verified evaluation results were extracted; KL divergence and refusal counts come from the publisher model card.
• Context length, layer counts, vision/video claims, thinking-mode defaults, and sampling presets come from the captured model card and have not been independently verified by Ethen for this heretic checkpoint.
• Provider state is historical snapshot data, not independently refreshed current availability.
• Qwen/Qwen3.8-27B lineage and Heretic conversion details come from the publisher model card, not an independently verified Ethen lineage judgement.

Source and provenance

Source: darkc0de/Qwen3.8-27B-heretic

Captured: Unknown. Processed: 2026-09-07T19:36:07.393415+00:00.

This is a decensored version of Qwen/Qwen3.8-27B , made using Heretic v1.4.0 This model is reproducible! See the README in the reproduce directory for more information. Abliteration parameters Parameter Value direction_index per layer attn.o_proj.max_weight 1.03 attn.o_proj.max_weight_position 62.77 attn.o_proj.min_weight 0.75 attn.o_proj.min_weight_distance 37.53 mlp.down_proj.max_weight 1.35 mlp.down_proj.max_weight_position 62.48 mlp.down_proj.min_weight 1.10 mlp.down_proj.min_weight_distance 19.33 Performance Metric This model Original model ( Qwen/Qwen3.8-27B ) KL divergence 0.0095 0 (by definition) Refusals 64/100 89/100 Qwen3…