Qwen3.8-27B-abliterated-GGUF

Qwen3.8-27B-abliterated-GGUF

Qwen3.8-27B-abliterated-GGUF is a GGUF-quantized conversion of huihui-ai/Huihui-Qwen3.8-27B-abliterated, an uncensored variant of Qwen/Qwen3.8-27B produced through abliteration — a crude, proof-of-concept activation-editing technique that removes refusal behavior directly from model weights without relying on TransformerLens. The underlying Qwen3.8-27B is a 27-billion-parameter dense causal language model with a native vision encoder, built on the Qwen3.5 architectural foundation, featuring a 64-layer hybrid design interleaving Gated DeltaNet linear-attention blocks with periodic Gated Attention layers, Multi-Token Prediction (MTP) training, a native 262,144-token context window extensible to 1M via YaRN, native image/video understanding, and flexible thinking control through a reasoning_effort parameter, delivering strong results on benchmarks like SWE-bench Pro (61.7), OSWorld-Verified (84.3), and GPQA Diamond (89.2). This GGUF release packages the abliterated weights across the standard quantization sweep for efficient local deployment via llama.cpp and compatible runtimes; note that, as with GGUF conversions generally, the Multi-Token Prediction (MTP) heads are not preserved in this format — the model runs as a standard single-token-per-step autoregressive decoder, so any latency or quality benefits tied to MTP-based speculative decoding in the original checkpoint do not carry over to these quantized builds.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp