Chinese AI Kimi K3 Matches US Rivals in GPU Code Speed

A Chinese AI startup has made a major leap in code generation. Moonshot AI's Kimi K3 model generated an NVIDIA H100 CUDA kernel that runs nearly 15 times faster than optimized PyTorch code.

The 2.8-trillion-parameter, open-weight model was released on July 16, 2026. It supports a one-million-token context window, allowing it to process massive codebases. Its architecture includes Kimi Delta Attention and a high-sparsity Mixture of Experts approach to manage computational costs.

In performance tests on NVIDIA hardware, Kimi K3 rivaled leading American models like Anthropic's Claude Fable 5 and came close to the performance of Claude Opus 4.8. It also produced a GPU compiler called MiniTriton, which reportedly matches or exceeds existing tools like torch.compile.

Moonshot AI plans to release the full model weights on July 27, making it one of the largest open-weight models ever published. The company's API pricing starts at $0.30 per million tokens for cached input.