Kimi K3 AI Model Optimizes Its Own Successor Using Self-Generated Code

Moonshot AI's Kimi K3, a 2.8 trillion-parameter frontier language model, was publicly released and topped the Code Arena frontend benchmark, surpassing Claude. A blog post dated July 16, 2026 revealed that an early version of Kimi K3 was used to handle most of the optimization work for later versions of the model itself. In four documented experiments, K3 achieved up to 2.48x faster training for AttnRes, a 55.1% reduction in DSA kernel runtime, and a 3.79x speed improvement on third-party GPGPUs. The model also independently built a small Triton-like compiler end-to-end and produced a near-manufacturable ASIC chip design as a proof of concept. Notably, K3 completed a complex computational research workflow — including reviewing over 20 papers and evaluating 300+ equations — in roughly two hours, a task that typically takes a skilled researcher one to two weeks.