Kimi in one minute
Kimi can answer questions, investigate documents and work with search and programming tools. On 8 October 2026, Kimi K3 is the reference family in the official product documentation. Long context helps with large dossiers, but does not replace source checking or a clearly defined task.
Kimi K3 and the assistant
Moonshot introduced Kimi K3 on 16 July 2026. K3 processes text and visual information, including images and video. The developer targets research, programming and tasks with multiple steps. Kimi K3 also has downloadable weights under its own Kimi K3 licence; do not automatically treat it as MIT or Apache 2.0.
Kimi on the web or in an app is a product with features, limits and subscriptions. Kimi Code focuses on software work. The API supplies models to developers. The same family name does not mean every access method offers identical context, tools or usage limits.
The product documentation distinguishes modes including K3 and K3 Cluster. Creating files, searching or running a program happens through the associated working environment. Isolated model weights do not perform these actions without supporting tools.
What is a large context useful for?
The general K3 model page lists a context window of one million tokens. Kimi Code includes k3 and k3-256k configurations. Check the access method and configuration you choose, rather than relying only on the largest number on a product page.
For example, an agency wants to identify differences between its existing product range, supplier agreements and a new campaign plan. A large context lets it supply much of that material together. Ask for the document and relevant passage behind each conclusion, and keep contradictions visible.
A million tokens is not a fixed number of PDF pages. Language, tables, images and conversion affect input size. A large context window is not perfect memory either: a detail can be missed or attributed to the wrong source.
Search, programming and actions
Kimi can use tools to retrieve additional information and carry out research in steps. In coding, an agent can edit files and call tests when the environment permits it. Define a clear result, such as a checked comparison or a proposed repository change.
A useful marketing process starts with source material and ends with a draft that can be checked. For example, research common product questions, group them by topic and prepare an FAQ. Check freshness and provenance before publishing. Moonshot’s research into long context and agents explains its technical direction, but vendor benchmarks do not establish that every research task will succeed autonomously.