Huihui-Qwen3.8-27B-abliterated

Upstream

Qwen3.8-27B

Modified by

huihui-ai

Method

Sumandora/remove-refusals-with-transformers

Canonical modified weights

available upstream (high-precision weights published)

Method (as documented): Sumandora/remove-refusals-with-transformers Method parameters: PARTIALLY PUBLISHED and usefully so: layers 18-51 ablated; the first 15 layers and other components (MTP and visual modules) remain unablated to preserve performance. No direction count, no scaling factors. Modifier: huihui-ai Notes and caveats: - base_model_import_note: imported during this run - base_model_note: Qwen/Qwen3.8-27B VERIFIED to exist, exact spelling, no -Instruct or -Base suffix. It is MULTIMODAL (pipeline_tag image-text-to-text), so every derivative of it is a vision model. - gguf_detail: 2682158 dl/30d, 711 likes. Ships mmproj-model-bf16.gguf (vision projector present), full Q2_K-Q8_0_L ladder, unsloth UD/UD-DW dynamic quants, GSQ-RCO variants, -bf16.gguf. - gguf_metadata_caveat: The GGUF repo declares base_model as Qwen/Qwen3.8-27B (tag base_model:quantized:) — it points at the UPSTREAM, not at the abliterated safetensors repo. The lineage link exists only in the naming, not the metadata. - adoption_note: The GGUF repo carries ~68x the traffic of the safetensors source. - evaluation_note: NONE, author or independent. Source: refusal-modified-acquisition-manifest.json v1.0.0 (research snapshot 2026-09-20), candidate 1.

Explore further