Chinese startup Moonshot AI has officially introduced the new Kimi K3 neural network. This open model has a whopping 2.8 trillion parameters. The system supports native vision and processes a context up to a million tokens long. The architecture is based on the innovative technologies of Kimi Delta Attention and Attention Residuals. The first mechanism efficiently processes long sequences of data. The second algorithm selectively extracts the necessary information from different layers. The Stable LatentMoE framework is responsible for computational efficiency. Of the 896 experts, only sixteen modules are activated at the same time. This approach resulted in a two-and-a-half-fold increase in productivity. The developers note that their models held the leadership in size among open systems.

The test results showed impressive performance in complex engineering tasks. Kimi K3 surpassed its proprietary counterparts Claude Fable 5 and GPT-5.6 Sol in the SWE Marathon and Program Bench benchmarks. The model is capable of conducting long-term engineering sessions with virtually no human involvement. She confidently navigates large repositories and manages terminal tools. In the GPU core optimization test, the system worked offline for up to twenty-four hours. In the late stages of development, the early version of K3 performed most of the optimization on its own. The neural network has also written a compact MiniTriton compiler for GPUs from scratch. This tool includes its own intermediate representation layer and code generation.

The model's capabilities go far beyond standard programming. Kimi K3 independently designed a chip for a neural network in just forty-eight hours. In the simulation, the device outputs more than 8,700 tokens per second at a frequency of 100 MHz. The system also successfully reproduced complex astrophysical relationships in just two hours. To do this, she analyzed hundreds of scientific articles and wrote thousands of lines of code. Separately, it is worth noting the native work with videos to create interactive prototypes. However, the developers honestly point out the current limitations of the system. The model is sensitive to loss of history when switching the agent environment. In addition, the usability is still inferior to the leading proprietary competitors.