Moonshot AI has released Kimi K3, a 2.8 trillion parameter open-weight model that generates CUDA kernels for Nvidia H100 GPUs at speeds 14.82 times faster than standard PyTorch. The performance leap directly challenges top US AI labs and deepens the already intense US-China artificial intelligence competition.
What the speedup means
The 14.82x improvement over PyTorch is not a synthetic benchmark — it’s a real kernel-generation speed. That means Kimi K3 can write and optimize low-level GPU code far quicker than the widely used PyTorch framework. For developers running large-scale training or inference on H100 clusters, the difference translates into dramatically shorter wait times and lower compute costs. Moonshot AI claims the model is open-weight, so researchers outside the company can inspect and adapt the architecture.
Kimi K3 arrives as Washington tightens export controls on advanced chips to China. US labs like OpenAI, Google DeepMind, and Meta have long dominated the frontier of large language models and GPU optimization. Moonshot AI’s model shows that Chinese firms can still push hardware performance without access to the latest Nvidia hardware — the H100 is export-restricted but widely available in China through pre-ban stockpiles. The 2.8 trillion parameter count puts Kimi K3 among the largest open-weight models ever released, rivaling the scale of closed models from US leaders.
The open-weight factor
By making Kimi K3 open-weight, Moonshot AI invites global developers to build on its work. That could accelerate adoption in China and elsewhere, while US companies typically keep their most advanced models proprietary. The move also pressures US labs to either open their own models or risk falling behind in the ecosystem of third-party optimizations. Nvidia’s CUDA platform remains the dominant GPU programming environment, but a model that writes CUDA kernels faster than PyTorch could shift how developers approach performance tuning.
Moonshot AI did not disclose the training cost or the exact hardware used to train Kimi K3. The company also hasn’t said whether the model will be updated or if it plans to release a version optimized for Nvidia’s upcoming Blackwell architecture. For now, the model is available for download, and the first independent benchmarks are expected within weeks.
