Kimi K3 is a 2.8-trillion-parameter open-weight model that pairs a highly sparse Mixture-of-Experts design with two new architectural ingredients, Kimi Delta Attention and Attention Residuals, giving it a hybrid attention structure that scales to long contexts while keeping inference efficient. It was developed by Moonshot AI as what the team calls the world's first open 3T-class model, with native vision support baked into the architecture alongside text. The model is positioned for frontier work on long-horizon coding, knowledge tasks, and step-by-step reasoning, and Moonshot reports it outperformed other open systems in its evaluation suite even though its overall scores still sit below the strongest proprietary competitors.

In practice, Kimi K3 fits teams that need a very large context window together with multimodal understanding, since it natively handles text, image, and video inputs and can produce structured, tool-assisted outputs. Moonshot committed to releasing the full weights by late July 2026, and vLLM worked with Moonshot AI, NVIDIA, and AMD on day-zero open-source serving, including KDA-aware prefix caching, multimodal integration, and tool-calling parsers. That ecosystem push makes Kimi K3 a strong option for self-hosted research agents, long-document analysis, and coding assistants where a million-token memory and an open-weights license matter more than chasing the absolute top of the closed-model leaderboard.