We've been asked why Cline works better with open weights models like DeepSeek than other harnesses and want to share some detail (more in our blog below). Our internal evals show that many open weights models like DeepSeek, GLM, and Kimi are RL trained to spend more tokens verifying their work: run tests, check the build, re-read diffs before completing. We believe this is how they were able to close the gap with closed model performance. Meanwhile closed models and their harnesses tend to optimize for efficiency and fewer tokens (which makes sense when massively subsidizing subscription plans). So we've been intentional in building the Cline harness to let the model work however it was trained to work. At open weights pricing, the baked-in paranoia burning extra tokens verifying work gets significantly better results for less money. In our benchmark runs, we've seen ~20% gains over competing harnesses just from changes that let the model lean into its RL training.