Moonshot AI's 2.8-trillion-parameter open-weight model is challenging the best American labs at GPU kernel optimization, and it's doing it at a fraction of the cost.
A Chinese AI startup just dropped a model that writes GPU code faster than most engineers would believe possible. Kimi K3, built by Moonshot AI, generated an H100 CUDA kernel that runs 14.82 times faster than optimized PyTorch, nearly matching the performance of Anthropic’s Claude Opus 4.8 on the same task.
What Kimi K3 actually is #
Released on July 16, 2026, Kimi K3 is a 2.8-trillion-parameter open-weight multimodal model. It supports a 1 million token context window, meaning it can digest entire codebases or book-length documents in a single pass. It achieves this through two novel architectural designs: Kimi Delta Attention (KDA) and Attention Residuals, paired with a high Mixture of Experts (MoE) sparsity approach that keeps compute costs manageable despite the enormous parameter count.
Moonshot AI plans to release the full model weights on July 27, 2026, making Kimi K3 one of the largest open-weight models ever published.
The CUDA kernel benchmark that matters #
Kimi K3’s performance on NVIDIA H200 hardware was particularly notable, rivaling leading American models like Claude Fable 5 and coming within striking distance of Claude Opus 4.8. The model also produced something called MiniTriton, a GPU compiler that reportedly matches or exceeds the capabilities of torch.compile and Triton, two widely used tools in the deep learning infrastructure stack.
The pricing play #
Kimi K3’s API pricing is structured at $0.30 per million tokens for cache-hit input, $3 per million tokens for cache-miss input, and $15 per million tokens for output.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our