Moonshot AI launches the long-context Kimi chatbot

Background

Kimi was released by Moonshot AI, a company founded in March 2023 — only seven months before the chatbot appeared.^[2]

The launch

In October 2023 Moonshot AI officially released the Kimi chatbot, at first in closed beta.^[1] The selling point from the beginning was how much text it could hold: Kimi handled up to 200,000 Chinese characters in a conversation.^[2]

On 16 November 2023 the chatbot opened to the general public, with 128,000 tokens of lossless context.^[1] The two figures are counted in different units — characters in the first case, tokens in the second.

Long context as the signature

Long context became Kimi's signature feature.^[1] In March 2024 the ceiling was extended to 2 million characters.^[1] The Star, in a report on 21 May 2024, described the chatbot as having been launched the previous October and noted that in March it could process up to two million Chinese characters in a single prompt, up from 200,000.^[3]

What a context window is

A large language model works inside a context window: the stretch of text it can consider at one time. Everything the model is meant to work from — the conversation so far, a pasted document, a set of instructions — has to fit inside that window. When the text runs past the limit, the system has to drop earlier material or replace it with a summary, and whatever has been dropped can no longer influence the answer. The word "lossless" points at the same idea: the input is meant to stay available in full rather than being compressed away before the model reads it.

Text is counted in tokens, which are roughly word-pieces rather than whole words. Because Chinese characters and tokens do not map one to one, the same capacity can be quoted as a token count or as a character count, and the numbers will differ.

Why the length mattered

For a user, the size of the window decides what kind of task is practical. A short window is enough for a brief exchange; a very large one is what makes it possible to hand over a long document in one go and ask questions about the whole of it, instead of cutting it into pieces and feeding them in one at a time. Fitting a whole text in at once also avoids the failure mode where the model answers a later question while having lost the earlier part of the material. Making the window itself the headline — rather than only the model's reasoning or writing quality — turns "how much can it read at once" into a product number a user can compare directly.

What the launch showed

Kimi's release schedule was itself part of the story. A company founded in March 2023 shipped a product in October of the same year and opened it to everyone in November.^[1][2] Between the launch and March 2024 the advertised capacity grew from 200,000 Chinese characters to 2 million, an increase of tenfold in a matter of months.^[1][2] By 21 May 2024, when The Star wrote about the service, long text was still the trait used to identify it.^[3]