All of the OS models I am trying to use via DeepInfra fail to execute plugins at varying rates. Every single time I request a plugin to be used, Mistral 7B just returns the function call string in the chat response, rather than actually using it to call the function. This happens less so, but still too frequently, with DeepSeek R1, Qwen3, and Llama Scout 4. This really debilitates the functionality of these models. Please fix!