Qwen 3.5 9B uncensored writer fine-tunes land in GGUF quantizations

Two uncensored fine-tunes of Qwen 3.5 9B optimized for creative writing appeared on HuggingFace on July 11 in GGUF format. Nuofang published quantized versions of davidau's F451-AND-TRI-Polar-Ultra-Pro-Writer-Uncensored-Heretic and The-Bradbury-F451-Pro-Writer-Uncensored-Heretic checkpoints, both built for local inference through llama.cpp. The F451 naming references Ray Bradbury's Fahrenheit 451, signaling the models' focus on unrestricted prose generation.

Both releases ship as imatrix-quantized GGUF files, a format that compresses the 9-billion-parameter base into smaller weight files suitable for consumer hardware while preserving output quality through importance-matrix calibration. The quantizations support llama.cpp and compatible runtimes, making them immediately runnable on CPUs or single-GPU setups without the memory overhead of full-precision weights.

Quantization and deployment

The underlying davidau fine-tunes start from Qwen 3.5 9B, Alibaba's instruction-tuned language model. Davidau's training removed safety alignment layers and added creative-writing datasets, producing checkpoints marketed as "heretic" variants—a term common in the uncensored model community for weights that bypass refusal behavior. The F451-AND-TRI-Polar variant combines multiple writing-focused datasets; The-Bradbury variant emphasizes narrative prose.

Neither model card includes benchmark scores, sample outputs, or training dataset specifics beyond the "uncensored" and "pro writer" labels. Both sat at zero downloads and zero likes at publication, typical for fresh GGUF quantizations before community testing surfaces quality reports. Practitioners looking for unrestricted creative-writing models in the 9B parameter range now have two options to test locally.