llama-server config from the video (8GB VRAM + 12GB+ system RAM)

llama-server \
  --host 0.0.0.0 --port 8091 \
  -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-IQ3_XXS \
  -c 131072 -ngl 99 --n-cpu-moe 25 \
  -b 2048 -ub 2048 \
  --cache-type-k q4_0 --cache-type-v q4_0 \
  --cache-ram 0 --parallel 1 --jinja \
  --reasoning on --reasoning-budget 8192 \
  --temp 0.7 --top-k 20 --top-p 0.8 --min-p 0.0 \
  --presence-penalty 0.0 --repeat-penalty 1.0 \
  --n-predict 32768

Same thing on one line:

llama-server --host 0.0.0.0 --port 8091 -hf unsloth/Qwen3.6-35B-A3B-GGUF:UD-IQ3_XXS -c 131072 -ngl 99 --n-cpu-moe 25 -b 2048 -ub 2048 --cache-type-k q4_0 --cache-type-v q4_0 --cache-ram 0 --parallel 1 --jinja --reasoning on --reasoning-budget 8192 --temp 0.7 --top-k 20 --top-p 0.8 --min-p 0.0 --presence-penalty 0.0 --repeat-penalty 1.0 --n-predict 32768

Notes:
- -hf downloads the GGUF from Hugging Face on first run (about 13 GB). It also pulls the mmproj vision tower; add --no-mmproj if you don't want vision.
- --n-cpu-moe 25 is for a dedicated inference box. If the same GPU drives your display, use 23.
- -c 131072 is 128k context with Q4 KV cache. If you run out of VRAM, drop to 65536 first.