Based on my tests if you are a personal AI person with 24GB Vram, best quant is not Q4_K_M versions. Q4_K_L quantization is offers better performance-size-speed trio. I think this model is performing better among other Qwen3.6 27B versions for general purpose agentic AI use case.

Original Model

Base: ManniX-ITA/Qwen3.6-27B-Omnimerge-v4
Architecture: qwen3_5
Parameters: 27B
Task: Image-Text-to-Text

Quantization

Format: GGUF
Quantization: Q4_K_L
Quantized by: kdqemre

Usage with llama.cpp (Optimized Params for 24GB VRAM- You can try longer context length)

bash

llama-server -m Qwen3.6-27B-Omnimerge-v4-MTP-Q4_K_L.gguf \
  -ngl 99 \
  -c 16384 \
  -fa 1 \
  -np 1 --cache-type-k q8_0 --cache-type-v q5_1 --reasoning-format deepseek --reasoning-budget 2048 \
  --spec-type draft-mtp --spec-draft-n-max 3 --cache-type-k-draft q5_1 --cache-type-v-draft q5_1 --kv-unified \
  --port 8088

License

Original model licensed under Apache-2.0. This quantized version preserves the original license.

kdqemre/Qwen3.6-27B-Omnimerge-v4-MTP-Q4_K_L-GGUF

Qwen3.6-27B-Omnimerge-v4-MTP-Q4_K_L-GGUF

Why this quantization exists?

Original Model

Quantization

Usage with llama.cpp (Optimized Params for 24GB VRAM- You can try longer context length)

License

kdqemre/
Qwen3.6-27B-Omnimerge-v4-MTP-Q4_K_L-GGUF