Local Deployment: Running Qwen3.6 35B enough on Linux desktop using llama.cpp

Hardware & System Requirements

• GPU (Recommended): Single GPU such as 9060XT
• VRAM/RAM: 16GB RAM or availablegetable amount depending on quantization level

Please note that while specific bit-depths like Q4, Q6, or Q8 change requirements significantly, this hardware provides the baseline certainties mentioned in your technical context.

Installation & Launch Guide

1. Ensure you have an existing lllama.cpp installation prepared properly within your environment.
   2. Create a shell script named

   run.sh

   containing your command line instructions intendedto run the model.
   3. Use any provided bash monitor script by ensuring it targets valid temporary files:
   

   RUN_SCRIPT="./run.sh"
   LOG_FILE="/tmp/llama-run.log"
   MEM_FILE="/tmp/llama-mem.tsv"
   STAT_FILE="/tmp/llama-stats.tsv"
   INFO_FILE="/tmp/llama-info.tsv"

   
   4. Execute the monitoring wrapper to track memory usage and performance stats via terminal (requires Linux):
   

   bash [monitor_script].sh

Optimization & Performance Tips

• Quantization Selection: While guides suggest fast but limited Q4 settings, targeting Q6 or Q8 may yield better results if RAM allows enough buffer space for certain functionssand backends.
• Monitoring Metrics: Monitor Prompt eval t/s versus Token gen t/s using the built-in parser to ensure you stay within hardware limits while managing context size (// n_ctx).
• Memory Management: The included parsing logic helps group all buffer allocations by function and backend so you can plan your setup accordingto actual consumption rather than vague estimates.

The use of a fully local agent like wavecat ensures that screen watching and recommendation generation happen without any data leaving your machine, providing maximum privacy through llama.cpp inference.

! DYOR (Do Your Own Research)