Moonshot AI has launched Kimi K3, an open-weight model with 2.8 trillion parameters—the largest released in its category. Its mixture-of-experts architecture contains 896 internal experts but activates only 16 per token, significantly reducing resource consumption and making it more efficient than dense models of comparable scale.
The model supports a 1-million-token context window, enabling it to process entire codebases, lengthy documents, and complete software projects in a single pass. It handles text, images, and video. Kimi K3 is available now through Moonshot's app and platform, with full model weights releasing July 27 under a permissive license. Pricing starts at $3 per million input tokens and $15 per million output tokens.
Performance benchmarks place Kimi K3 at 57 points on Artificial Analysis's intelligence index—close to Anthropic's Opus 4.8 but behind Fable 5 and GPT-5.6 Sol. On practical task evaluations, it ranks third and outperforms Opus 4.8. For developers, the standout result came in Arena rankings, where Kimi K3 led in web interface programming, beating established American models.
The open-weight distribution, flexible licensing, and competitive performance create opportunities for customization, local infrastructure deployment, and AI agent development without full reliance on proprietary APIs. However, Moonshot disclosed many results internally, independent testing remains limited, and some evaluators reported higher error rates compared to the previous generation.
