Kimi K3 is a 2.8-trillion-parameter model built around Moonshot's Kimi Delta Attention and Attention Residuals architecture, with native vision capabilities and a 1-million-token context window. It is positioned as the first open 3T-class model, and its reasoning depth can be dialed up or down by configuring reasoning effort (with a max default). It is aimed squarely at long-horizon coding and end-to-end knowledge work, including workflows that rely on tool calling, structured JSON output, and automatic context caching to amortize repeated prompt content.

On third-party leaderboards, K3 places second on the Vals AI index and reaches an Artificial Analysis Elo of 1547 on long-horizon knowledge work, a step widely described as bringing the open-to-closed gap down to roughly three to five months while still trailing the top US proprietary systems. The Simon Willison newsletter notes open weights were promised by late July, and Synthetic now lists K3 alongside other open-source families it serves for private inference. That combination of open weights, a very large context window, and frontier-adjacent reasoning makes K3 a strong fit for teams experimenting with agentic coding, multi-document research, and tool-augmented workflows who want flexibility without committing to a closed frontier API.