Qwen3.6 27B uncensored abliterated i1 IQ4 XS GGUF(Smaller) This model is equivalent to https://huggingface.co/lemonyins/Qwen3.6 27B abliterated i1 IQ4 XS GGUF Smaller Innovation This model refers to the fully optimized Qwen3.6 27B i1 IQ4 XS, which has restored the attn qkv layer to pure IQ4 XS . Furthermore, we introduce a novel hybrid precision quantization strategy: the FFN layer uses IQ3 S , which achieves significantly smaller file sizes while maintaining core inference capabilities through support from TurboQuant KV caching. Additionally, we use an Huihui abliterated version of the base model for quantization, making it convenient for users to conduct in depth research. Motivation The original llama.cpp quantization heuristics upgrade attn qkv layers to q5 K under certain conditions (e.g., n gqa = 4 ), causing noticeable file bloat. cHunter789's fix restores attn qkv to pure IQ4 XS , saving ~375 MiB. Taking this further: FFN layers (ffn down, ffn up, ffn gate) account for ~2/3 of total model parameters , yet they have higher redundancy than attention layers. Downgrading them from IQ4 XS to IQ3 S is a natural next step — it yields substantial size reduction with minimal quality…
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy