Moonshot AI built this model as a coding-focused successor to K2.6, reshaping the earlier foundation into an agentic tool aimed at long-horizon software engineering. It is released as an open-weights model under a Modified MIT license, with weights published on Hugging Face in mid-June 2026, and the free routing path through ZenMux lets developers experiment without per-token charges. The architecture is a one-trillion-parameter Mixture-of-Experts design that activates roughly 32B parameters per token, drawing on a large pool of specialized experts to keep each forward pass efficient. Compared with its predecessor, Moonshot trimmed thinking-token usage by around thirty percent, so multi-step coding jobs consume fewer reasoning tokens for the same outcome.

In practical terms, the model is shaped for finishing multi-step coding tasks end to end rather than answering isolated prompts, with thinking mode kept on across turns and native INT4 quantization to ease local deployment. It handles text and image inputs alongside experimental video, and it operates inside a generous context window that supports lengthy codebases and extended agent sessions. On Moonshot's reported coding suites it climbs meaningfully over K2.6, including a jump from 50.9 to 62.0 on Kimi Code Bench v2 and from 72.8 to 81.1 on MCPMark Verified, signaling stronger agentic tool use even where frontier closed models still lead. For developers, this combination of open weights, a permissive license, and a free hosted route through ZenMux makes it a pragmatic fit for self-hosted agent workflows and rapid prototyping of coding assistants.