When tackling the more complex Tetris game, however, none of the models were able to produce a fully working implementation. Deepseek-Coder-V2-Lite-Instruct generated some code that ran but had significant bugs, while Yi-Coder-9B-Chat and Qwen2.5-Coder-7B-Instruct both failed to create a functional Tetris game.

The real test came with more challenging algorithmic coding problems from the popular Code Wars platform. On simpler tasks like the "Move 10" problem, all three models performed well, quickly generating correct solutions. But when faced with tougher challenges that required deeper problem-solving skills, the limitations of these offline coding assistants became apparent.