Debugging is a developer's daily grind—often frustrating and time-consuming. Recent advancements in large language models (LLMs) like ChatGPT have hinted at their potential as powerful debugging assistants. However, sharing sensitive code with external LLMs raises privacy concerns for many companies. Could open-source LLMs be a viable alternative? A new study evaluates the debugging prowess of several open-source LLMs on a benchmark called DebugBench, featuring over 4,000 buggy code examples in Python, Java, and C++. The results are promising, with DeepSeek-Coder leading the pack, achieving a 66.6% success rate in fixing bugs across all languages. While not yet as powerful as closed-source giants like GPT-4, these open-source models offer a compelling balance between performance and privacy. Interestingly, model size wasn't the sole determinant of performance. Smaller models sometimes outperformed their larger counterparts, indicating efficient fine-tuning and training data play crucial roles. The study also explored the link between coding ability and debugging skills. While a correlation exists, it wasn't absolute. Some models excelled at coding but lagged in debugging, suggesting these two skills might require different approaches when training LLMs. Challenges like data contamination—where training data overlaps with evaluation data—remain a concern. Despite using DebugBench, published after the models' knowledge cut-off dates, the possibility of pre-training contamination from LeetCode code still exists. Further research will investigate diverse code types beyond algorithmic problems and examine how techniques like prompt engineering can boost the performance of open-source LLMs for real-world, complex debugging tasks. The future of debugging might just be open source after all.