Qwen3.6-27B IQ4_XS FULL VRAM with 110k context
Qwen3.6-27B IQ4_XS Bloat: Reverting llama.cpp commit saves 16GB VRAM (14.7GB vs 15.1GB) + KVCache Tests With the release of Qwen3.6-27B, I noticed that compared to the excellent IQ4_XS quantization (14.7GB) by mradermacher for the 3.5 version (Qwen3.5-…