The Kimi K3 model has officially launched as the world’s first open 3T-class artificial intelligence system, boasting 2.8 trillion parameters. Developed by Moonshot AI, this massive open-weights release introduces native vision capabilities and a 1-million-token context window. It represents a major step forward in open-source artificial intelligence, offering frontier-level performance across complex coding, reasoning, and knowledge tasks.
Architectural Innovations of the Kimi K3 model
The Kimi K3 model is built on two primary architectural updates: Kimi Delta Attention (KDA) and Attention Residuals (AttnRes). These technologies improve how information flows across long sequence lengths and deep model layers. Additionally, the system scales up Mixture of Experts (MoE) sparsity, activating 16 out of 896 experts. This structural design yields a 2.5 times improvement in scaling efficiency compared to its predecessor, Kimi K2.
Advanced Coding and Compiler Development
In software engineering, the model operates with minimal human oversight to sustain long coding sessions. During internal testing, it successfully built MiniTriton, a compact GPU compiler, from scratch. This compiler delivers performance that matches or exceeds existing industry standards on specific workloads. The model also excels at tasks that combine software development with visual reasoning, using screenshots to optimize user interfaces and 3D designs.
Interactive Knowledge Work and Visualization
For complex research, the model introduces new interactive features within software applications. Users can generate dynamic widgets and personalized dashboards directly inside their chat sessions. In one case study, the model analyzed 42 years of industry research by executing over 2,800 web searches and processing 11,000 pages of documentation to build an interactive visual website. It also supports advanced video editing, translating technical concepts into animated diagrams.
Pricing and Platform Availability
Developers can access the Kimi K3 model today through various official channels, including the Kimi API. The pricing structure is set at $0.30 per million tokens for cache-hit inputs, $3.00 per million tokens for cache-miss inputs, and $15.00 per million tokens for outputs. To achieve optimal performance, the developers recommend deploying the system on supernode hardware configurations with 64 or more accelerators. The full model weights are scheduled for public release by July 27, 2026, as detailed on the official Kimi blog.
