What Is Kimi? Models for Long Tasks and Legacy APIs
Kimi is Moonshot AI’s model and application brand. This article uses the current model catalog and K2.5 open weights to explain long-context uses, legacy API retirement and the limits of the modified MIT license.
Meet this model
Kimi is the AI model and application brand of Moonshot AI. Using the Kimi chat app and integrating a model through the API are different approaches; downloading an older open-weight release is another deployment choice. Long-context and tool-task capabilities must be interpreted for a specific version. Retirement of a hosted API does not automatically invalidate downloaded files, but neither does it mean the old model remains callable at its former endpoint.
Current model roles differ from the chat app
The official API catalog on October 20, 2026 lists kimi-k3, kimi-k2.7-code and kimi-k2.6. K3 supports native visual understanding and a one-million-token context; K2.7 Code targets coding tasks. These roles help narrow the candidates, but workspace features in the chat app are not automatically features of the model API.
For a long research report with charts, prepare text and image inputs and ask for page references supporting each conclusion. Context capacity permits more material but does not guarantee extraction of every exception. For a few pages of tables, correct document parsing can also affect the result, so window size alone is an insufficient comparison.
Separate hosted model retirement from weight releases
The official catalog states that kimi-k2.5 and the moonshot-v1 series were discontinued on August 31, 2026; the kimi-k2 series was discontinued on May 25. A working example in an older tutorial does not guarantee access today. Match the model name in the actual request to the current catalog rather than relying on a configuration label that merely says “supports Kimi.”
The official Kimi-K2.5 open-weight model card remains available. It describes a mixture-of-experts architecture with one trillion total parameters and 32 billion active per token. Active parameters describe part of the computation, not a requirement to store only 32 billion parameters. Do not estimate total self-hosting memory from the active count alone.
Evaluate licensing and tool execution separately
K2.5 weights use a modified MIT license. Its additional condition requires commercial products using the software or derivatives to prominently display “Kimi K2.5” in the user interface if monthly active users exceed 100 million or monthly revenue exceeds US$20 million. This is not a license statement for all Kimi versions, nor does the older weight license establish current API entitlements.
The API quickstart provides OpenAI-compatible and Anthropic-compatible access, while tool workflows still need correct application handling. Creating a spreadsheet from a report requires a real file-generation tool and its execution result; a filename mentioned in an answer is insufficient. This article reviews the official catalog and model cards without hands-on report processing or local deployment tests.
Checked sources
These sources support the concepts and methods described here. Examples and test scope are labeled separately; check the version of the tools you use.