My llama.cpp Settings for Qwen3.8 27B

Following up on trying different AI models, Qwen3.8 is the kind of model I was hoping for, something smart enough to be genuinely useful that can still run on my hardware. So I've been running it locally with llama.cpp. Sharing my setup below.

My current machine is an AMD Ryzen 9 5900X, 32GB of RAM and an ASUS 5060 with 16GB of VRAM. A lot more to work with than my old setup, but here's how I launch it:

llama-server.exe ^
  -m "models\Qwen3.8-27B-Uncensored-IQ4_XS_4BPW.gguf" ^
  --host 0.0.0.0 ^
  --port 8080 ^
  -c 62800 ^
  --parallel 1 ^
  -fa on ^
  --fit on ^
  --cache-type-k q4_0 ^
  --cache-type-v q4_0 ^
  --reasoning-budget -1 ^
  --chat-template-kwargs "{\"preserve_thinking\":true}" ^
  -a model ^
  --jinja ^
  --chat-template-file "templates\chat_template.jinja" ^
  --reasoning-format deepseek ^
  --reasoning-preserve ^
  --kv-unified ^
  --no-cache-idle-slots

I've not explicitly benched it, but on this setup it's around 25 tokens a second, which is plenty for what I use it for.

Aidan.