Worked on enhancing the GGUF export process in the unsloth-zoo repository, focusing on improving reliability and compatibility for MTP-enabled models. Addressed a critical bug by stripping unnecessary MTP-specific configuration keys and restoring architecture information from the original Hugging Face config, ensuring accurate model conversion. Used Python for backend development and data processing, integrating closely with the llama.cpp converter. The solution prevented block count inflation during export, reduced errors across Studio UI, API, and internal converters, and streamlined the end-to-end workflow. This work enabled more robust exports, supporting machine learning engineers with faster, more reliable model deployment and integration.