Qwen2.5 instruct models, self-quantized to GGUF by Atomic Chat with a per-tensor importance matrix.