Kimi is a generative artificial intelligence assistant developed by Moonshot AI, a company based in China. Launched in October 2023, it provides a conversational interface for answering questions, analyzing documents, drafting text, and assisting with programming. Long-context processing—working with substantial amounts of text within a single interaction—was a central feature of its initial positioning. The name Kimi is also used for a family of models developed by Moonshot AI; these models and the consumer chatbot are related but are not identical products.
Development and naming
Kimi emerged during the expansion of consumer services based on large language models. Its early emphasis was on reading lengthy Chinese and English material, including uploaded documents, and answering questions about that material. Moonshot AI subsequently expanded the service with web search, additional document-processing functions, and reasoning-oriented capabilities.
The Kimi name has increasingly served both as a product brand and as a model-family name. In January 2025, Moonshot AI published Kimi k1.5, a model described in its technical report as using reinforcement learning to improve reasoning across text and visual inputs. In July 2025, it released Kimi K2, a mixture-of-experts language model designed for tasks including coding and tool use.
These releases are milestones in the underlying model program, rather than evidence that every version of the chatbot exposes the same capabilities. The hosted assistant, developer services, and downloadable model releases can differ in their available tools, input limits, and deployment conditions.
Functions
Conversation and writing
Kimi accepts natural-language instructions and produces responses using the current conversation as context. Common functions include explanation, brainstorming, rewriting, translation, and drafting. Its writing capabilities range from short answers to structured documents, although following a requested format does not guarantee factual accuracy.
Programming assistance can include generating code, explaining existing programs, and suggesting changes. Producing code is distinct from executing or testing it: whether the assistant can run a program depends on the tools available in the particular product mode.
Document analysis and long context
A distinguishing feature of Kimi is its support for lengthy inputs. Users can supply documents and ask for summaries, comparisons, extracted information, or answers grounded in the supplied text.
The relevant technical concept is the context window: the amount of information a model can consider during a generation. This includes instructions, conversation history, supplied material, and space reserved for the response. Models process text through tokenization, so a limit expressed in tokens is not directly equivalent to a fixed number of words, characters, or pages.
A larger context window permits more material to be supplied at once, but does not ensure that every passage receives equal attention. Performance can vary with document structure, the location of relevant information, and the number of reasoning steps needed. Extracting text from complex tables, scanned pages, or unusual layouts also introduces a separate source of error.
Web-assisted answers and tool use
Kimi offers web-assisted answering, combining conversational generation with information retrieval. Retrieved pages can provide information that was unavailable when the underlying model was trained. This broadly follows the principle of retrieval-augmented generation, in which external material is supplied as context for an answer.
Some capabilities also involve tool use: the model selects an operation, supplies inputs, and incorporates the resulting output into its response. Such workflows can support AI-agent behavior when they extend beyond a single answer into a sequence of actions. Tool availability depends on the interface and service configuration.
Technical basis and access
Kimi belongs to the class of assistants built around neural language models, generally using the transformer architecture. These models learn statistical patterns from training data and generate responses conditioned on an input sequence. Additional training can improve instruction following, reasoning, and tool use.
The consumer service packages model capabilities with an interface, document handling, retrieval systems, and other software. This surrounding infrastructure matters: a chatbot's behavior cannot be inferred from the underlying model alone.
Moonshot AI also provides developer access through an application programming interface. An API allows software to submit requests and receive model outputs without using the consumer interface. Separately released model weights permit other forms of deployment, subject to their individual licenses. Availability of weights does not by itself establish that a release meets every definition of open-source software, nor does it imply publication of all training data or training code.
Limitations and evaluation
Like other generative assistants, Kimi can produce hallucinations: fluent statements, quotations, or explanations that are unsupported or incorrect. Document access and web retrieval can reduce some information gaps, but neither guarantees faithful interpretation or accurate citation.
Long-context performance requires evaluation beyond the advertised maximum input length. Relevant tests include locating information, combining evidence across passages, distinguishing conflicting accounts, and maintaining accuracy as the input grows. Likewise, results for a named Kimi model do not automatically describe the hosted chatbot, which may use different prompts, tools, or model configurations.
Processing uploaded material also raises data privacy considerations. Data handling depends on the particular service, deployment arrangement, and applicable terms. External documents and web pages can additionally contain prompt injection attempts—text intended to redirect the assistant rather than merely supply information—making the separation of user instructions from untrusted content an important technical concern.