Moonshot AI

Kimi runs on Moonshot AI's own model line, and the current flagship is Kimi K3. The company describes it as a 2.8 trillion parameter model with a new architecture, native multimodality and a context window of one million tokens. It's pitched at three jobs in particular: long-horizon coding, knowledge work and deep reasoning. That context figure is the important part. A million tokens means you can, in principle, drop in a large codebase, a stack of reports or a long research thread and keep it all in view instead of chunking it by hand.

On top of the model, Moonshot layers agent tooling. The site lists an AI Agent mode and an "Agent cluster" (Agent 集群), plus task-specific modes for deep research, one-click website building, smart spreadsheet creation and autonomous PPT editing. There's also Kimi Code for programming work, a Kimi Open Platform for API access, a Kimi Enterprise offering and something called Kimi Claw, which the homepage names but doesn't explain. You can reach all of it through a mobile app, a browser plugin, or desktop clients for macOS and Windows.

The release cadence is quick. Kimi K2.6 shipped on 2026-04-20 and K3 followed on 2026-07-16, the same day Moonshot published a benchmark called PerceptionBench. The research team says it's working toward AGI and shares work with the open source community, which explains why Kimi gets attention from developers who'd otherwise default to Western models. The catch for many readers: moonshot.cn is written in Chinese, and the homepage is long on positioning and short on specifics like usage limits or K3 benchmark numbers. If you can work with that, it's a free assistant with an unusually large context window and a broad set of agent modes.