v1k

llama-server by default in most implementation keeps the reasoning content in reasoning_content variable in response attribute. You can get it from there. Otherwise use reasoning-format flag and pass DeepSeek value to get pure tokens

This is good addition, Thank you.