We don't support Codestral or Qwen2.5-coder right out of the box for now, but depending on your use-case we certainly could add it. We utilize LoRA for smaller

We don't support Codestral or Qwen2.5-coder right out of the box for now, but depending on your use-case we certainly could add it. We utilize LoRA for smaller models, and qLoRA (quantized) for 70b+ models to improve training speeds, so when downloading model weights, what you get is the weights & adapter_config.json. Should work with llama.cpp!