Qwen3.8-27B-IQ5_KS-GGUF

Qwen3.8-27B-IQ5_KS-GGUF

MTP capable and smaller non MTP IQ5_KS quants of Qwen 3.8 27B for use with ik_llama.cpp, created through reuse of Ubergarm's tensor recipe for Qwen 3.6 27B with his updated Qwen 3.8 imatrix.

Creation of this was possible due to the more or less identical structure of 3.8 27B to its predecessor, and if the recipe is equally effective (which it should be) this should maintain SOTA performance for full 24GB offload with room for medium length full precision context windows.

Files

The no-MTP file is 430.413 MiB smaller. With MTP disabled, the tested ik_llama.cpp loader ignores the embedded predictor tensors, so both files use the same VRAM but 430MiB less disk usage is 430MiB less disk usage.

MTP usage

./llama-cli \
  -m Qwen3.8-27B-MTP-IQ5_KS.gguf \
  -ngl 999 \
  --spec-type mtp:n_max=1,p_min=0.0

Chat Template

Both GGUFs contain an exact match of the latest official Qwen3.8-27B chat template at the time of publication.

Sources and tooling

• BF16 GGUF Source: ggml-org/Qwen3.8-27B-GGUF
• Tensor Recipe: ubergarm/Qwen3.6-27B-GGUF
• Importance Matrix: ubergarm/Qwen3.8-27B-GGUF
• Individual-tensor quantizer: Thireus/ik_llama.cpp build 5263