Qwen3.6-27B with 100K Context on 16GB VRAM

To run Qwen3.6-27B on an A5000 16GB VRAM setup, a self-made IQ4_XS GGUF (qwen3.6-27b-IQ4_XS-pure.gguf) was used.

• The quantized build made with Unsloth imatrix was compared against other quants using mean KLD as the metric.
• It was noted that the buun-llama-cpp fork showed better results than the TheTom/llama-cpp-turboquant fork.
• The llama-server execution example used a context of 100,000 (-c 100000) along with options like --fit off and -fa on.
• Build commands for Windows and an example opencode.json configuration for OpenCode integration were also provided.

Since the hardware, model file, build method, and execution options needed for reproduction are all included, this serves as a useful reference case for running long-context local LLMs.