Hi everyone. I've been evaluating AI coding assistants for handling large codebases and long conversations, specifically for tasks like refactoring across multiple files or debugging complex, interconnected issues. My team currently uses a mix of tools, and I wanted to see how the much-discussed long-context capabilities of Kimi and DeepSeek compare in a practical coding scenario.
I designed a benchmark where I provided both assistants with a codebase containing 15 interconnected files (a mix of Python backend and React frontend, roughly 12k tokens total) and asked them to perform a specific refactoring: changing an API response structure and updating all dependent frontend components. I provided the entire context in a single prompt with clear instructions.
Here are my key observations:
* **Kimi** successfully parsed all files and identified the dependencies correctly. Its suggested refactor was logically sound, but the code changes were somewhat conservative and it missed updating two utility functions that were indirectly linked.
* **DeepSeek** also maintained coherence across the entire context. Interestingly, its approach was more aggressive; it refactored the core API file and all direct component dependencies, and it also proactively suggested a new utility function to consolidate repeated logic, which was a smart addition. However, it made one small syntax error in a React hook that I had to correct.
For pure accuracy in following instructions exactly, Kimi felt slightly safer. For a more insightful, "big picture" refactor that improves the code structure beyond the immediate request, DeepSeek was more impressive, though it required a bit more careful review of its output.
My main question for the community is about consistency. Has anyone else run similar long-context, multi-file coding tasks? I'm particularly curious about how these tools handle subsequent, follow-up questions within the same long conversation. Does one assistant tend to "forget" or contradict details more than the other as the chat extends?
Thanks!