Kimi K3 is an open-weight native multimodal model developed by Moonshot AI, designed as a frontier intelligence system for long-horizon coding, knowledge work, and reasoning. It introduces a new architecture built on Kimi Delta Attention (KDA) paired with Attention Residuals (AttnRes), combined with a Stable LatentMoE framework that activates 16 of 896 experts. The total parameter count reaches 2.8 trillion, making it described as the world's first open 3T-class model, while maintaining native vision capabilities alongside text understanding and a one-million-token context window for handling extensive inputs.

According to Moonshot AI's published evaluation suite, Kimi K3 delivers frontier-level performance that consistently outperforms other tested open models, though its overall scores still trail the most powerful proprietary systems such as Claude Fable 5 and GPT 5.6 Sol. The combination of massive MoE sparsity, extended context handling, and multimodal input support positions it well for complex agentic workflows that require sustained reasoning over large documents, codebases, or mixed visual-textual information. Its open-weight availability under the Kimi K3 license makes these capabilities accessible to researchers and developers seeking to build on or fine-tune a frontier-scale model locally.