I concur with this wholeheartedly.

This is rocket surgery. Right now, correctly running local LLMs is not user friendly for people who genuinely aren't in deep.

If you are using Ollama, you're flying blind. You have no idea what heinous BS they are launching that model with...or whether they even are using model card recommendations or the right chat template...or anything really.

Since you're on Windows OP, I'll share my Windows based Qwen3.6-27B-UD_Q5_K_XL launch parameters in hopes it inspires you to go switch to llama.cpp and learn what each parameter means and how to set your own parameters to avoid issues just like this one:

$env:PATH = "C:\llama.cpp\release\b8981;C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin;" + $env:PATH $env:CUDA_PATH = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1"

& "C:\llama.cpp\release\b8981\llama-server.exe" `
  --model "C:\llama.cpp\models\Qwen3.6-27B-UD-Q5_K_XL.gguf" `
  --mmproj "C:\llama.cpp\models\Qwen-3.6-27B-mmproj-BF16.gguf" `
  --no-mmproj-offload `
  --spec-type ngram-mod --spec-ngram-mod-n-match 8 --spec-draft-n-min 4 --spec-draft-n-max 16 `
  --n-gpu-layers 999 `
  --ctx-size 196608 `
  --parallel 1 `
  --threads 16 `
  --temp 0.8 `
  --top-p 0.95 `
  --min-p 0.01 `
  --top-k 30 `
  --repeat-penalty 1.0 `
  --repeat-last-n 256 `
  --presence_penalty 0.0 `
  --chat-template-kwargs '{\"preserve_thinking\": true}' `
  --mlock `
  --flash-attn on `
  --cache-type-k bf16 `
  --cache-type-v bf16 `
  --kv-unified `
  --host 0.0.0.0 `
  --port 8083

Here's also a link to all of the parameters available in llama.cpp to learn: https://github.com/ggml-org/llama.cpp/blob/master/tools/server/README.md