Overview
- MoonshotAI released Kimi K3 with public model weights, a technical report and the supporting infra stack (MoonEP, FlashKDA and AgentEnv) on GitHub and Hugging Face for immediate community access.
- Kimi K3 uses Kimi Delta Attention and Attention Residuals and expands a Mixture‑of‑Experts setup that can pick 16 experts from 896 to improve how capacity is used.
- MoonshotAI says the architecture and training changes yield about a 2.5× scaling-efficiency gain versus its prior K2 model, meaning more of the same compute is converted into capability.
- The model claims native visual understanding, a 1,000,000‑token context window and stronger long-range coding and tool-using abilities for large codebases and extended engineering tasks.
- MoonshotAI positions Kimi K3 as the largest open-source model but concedes it still trails leading closed systems such as Claude Fable 5 and GPT-5.6 Sol, and the public release could speed research, downstream experiments and ecosystem competition.