Kimi extends its context window to 2 million Chinese characters

Background

Kimi launched in October 2023 and could then take in up to 200,000 Chinese characters.^[2][3] That number describes the model's context window: how much text it can keep in view while producing one answer. The size of the window decides what kind of job an assistant can do in a single pass. A question about a few pages fits easily; a whole contract, a novel, or a long technical manual does not, and the usual workaround is to cut the document into sections and feed them in one at a time, which makes it hard to connect a sentence near the start with a clause hundreds of pages later.

What changed

In March 2024 Moonshot AI said Kimi could now handle up to 2 million Chinese characters in a single prompt.^[3] Compared with the launch figure that is about a tenfold increase, and it moves the practical unit of work from a single article towards something closer to a book-length manuscript. A user no longer has to decide in advance which chapter matters.

A staged release

The larger window did not reach every user at once. In March 2024 it ran as a closed beta, limited to invited users before a wider roll-out.^[1][3] Staged releases are common for features whose load and cost are hard to predict in advance: a small group of testers shows how the system behaves when people paste in genuinely long material, and what has to be fixed before everyone else is let in.

The same quarter

In February 2024, the quarter in which the beta began, Moonshot raised a US$1 billion round led by Alibaba.^[2][4] Money and context length are connected. Serving very long inputs is compute-heavy, because every extra character in the prompt is more work for the model on each request; an assistant that invites users to paste in entire books carries a different cost profile from one that answers short questions.

Caching and users

In July 2024 context caching entered public beta.^[1] The idea is to process a long stretch of text once and keep the result, so that follow-up questions about the same document do not mean starting again from the beginning each time. It pairs naturally with a large window: the more material a user pastes in, the more there is to gain from not reprocessing it on every turn.

By August 2024 Kimi ranked third among chatbots by monthly active users.^[2]

Why it matters

Context length is one of the quieter constraints on how language models get used. A model with a short window forces the user to do the summarising and selecting first, and anything left out is invisible to the model. A long window lets the raw material go in as it is, which makes whole-document work possible: going through a full report, checking a long agreement against a list of points, or keeping a drawn-out correspondence consistent.

The limit is worth stating plainly too. A larger window is a ceiling on what can be supplied, not a promise about how well any particular passage will be used, and processing long inputs is expensive enough that techniques like caching become part of making the feature practical. The March 2024 beta was a step in that direction rather than the end of it.