KIMI-DEV-72B (HUGGING FACE REPO)

Summary from the TLDR AI newsletter

Moonshot AI's Kimi-Dev-72B achieves 60.4% on SWE-bench Verified—the new state-of-the-art among open models—by using large-scale reinforcement learning that patches real repositories in Docker and only rewards solutions when the entire test suite passes.