The Devin maker tuned Moonshot's open model with reinforcement learning and says the result lands within a point…