AI Briefing
KO

Qwen3.6-27B IQ4_XS Regression

·2026.04.28 21:59

Key point

A llama.cpp regression increased the Qwen3.6-27B IQ4_XS size from 14.7GB to 15.1GB.

Details

The Qwen3.6-27B IQ4_XS GGUF grew from 14.7GB in the Qwen3.5 version to 15.1GB, reducing headroom on 16GB VRAM setups.

  • The llama.cpp commit 1dab5f5a44, identified as the cause, forced the attn_qkv layer's quantization to a minimum of Q5_K.
  • The author reverted this change, restoring the original IQ4_XS layer configuration 1:1.
  • Comparison using mradermacher's imatrix showed no significant quality degradation.
  • A 65k context PPL test based on pg19.txt and KV cache experiments were also conducted.
  • The fixed version cHunter789/Qwen3.6-27B-i1-IQ4_XS-GGUF has been published on Hugging Face.
llama-cppqwen3performance-regressionkv-cacheiq4-xs
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

RedditBeeLlama.cpp v0.4.1 Released: Improved KV Cache Quantization PerformanceRedditCachyLLama Supports SSD-Based KV CachingHuggingFace BlogHugging Face Adds GGUF Model Support to TransformersKT CloudHierarchical Memory Architecture Needed to Resolve AI Inference BottlenecksPyTorchKRDeepSeek Releases V4.1-Flash: Replaces V4-Pro and Optimizes KV CacheHacker News'Cache-to-Cache' Framework Revealed for Direct KV-cache Communication Between LLMs Without Text