--- title: Kimi description: "Moonshot AI's model family, including the Kimi K2 and K3 releases, whose architecture, capabilities, and licenses must be identified by version." canonical_url: 'https://darkfactory.dev/glossary/kimi' markdown_url: 'https://darkfactory.dev/glossary/kimi.md' collection: glossary date_published: '2026-10-07T00:00:00-04:00' date_modified: '2026-10-07T00:00:00-04:00' --- # Kimi ## Definition Kimi is Moonshot AI's family of models for language, reasoning, coding, and, in supported versions, visual understanding. The name also appears on the company's applications. Identify the underlying model release separately from the application or agent harness using it. The family includes architecturally different releases with different inference contracts. A statement about Kimi K2, K2.6, or K3 should retain that version whenever its capabilities, costs, or licensing matter. ## Origin and attribution Moonshot AI develops the family, and its Kimi Team publishes the technical reports. The team's Kimi k1.5 report appeared in January 2025 and documented reinforcement-learning methods for a multimodal model. Its July 2025 Kimi K2 report introduced a mixture-of-experts model with downloadable base and post-trained checkpoints. The Kimi K3 report followed in July 2026. These dates identify documented releases in the family's development; they do not establish the first use of the Kimi product name. ## Family, versions, and capabilities K2 has one trillion total parameters and 32 billion active parameters. K3 has 2.8 trillion total parameters, 104 billion active parameters, native vision, and a one-million-token context window. Those are release-specific architecture and capacity descriptions, not a general definition of every Kimi model. The October 7, 2026 API catalog lists `kimi-k3`, `kimi-k2.7-code`, its high-speed variant, and `kimi-k2.6`. It marks older K2 and K2.5 API offerings as discontinued. An API's support lifecycle is separate from the availability of previously released weights. Thinking behavior also varies. Current K3 always reasons, with configurable effort. K2.6 allows thinking to be disabled; K2.7 Code keeps it enabled. Their requirements for returning prior reasoning content through the conversation need to be followed for the selected model. ## Licensing and limits The K2 repository uses a modified MIT license with an additional attribution condition for large commercial products. K3 uses the separate Kimi K3 License, which adds commercial conditions for some Model as a Service businesses and large products, with stated exceptions. Do not transfer one release's license to another because both have downloadable weights. Moonshot's K3 evaluation notes identify different agent harnesses, reasoning settings, and tool augmentation across comparisons. A benchmark result measures the tested configuration. It does not isolate the family name as the cause of performance or guarantee success on a new workflow. ## Operational significance Record the exact model, endpoint or weights, inference engine, reasoning settings, and tool permissions. For local deployment, active parameter count alone does not specify how much hardware is required. For API use, confirm current support and the selected model's conversation contract. Test useful work and failure cases under the intended harness. A long context window supplies capacity, while evidence coverage and correct tool actions remain things to measure. ## Distinguish it from nearby terms - A foundation model is a broad model category. Kimi names a provider's family. - A mixture of experts is an architecture used by K2 and K3. It does not identify the provider or every family variant. - A reasoning model uses inference effort for more involved problems. Thinking controls differ across Kimi releases. - A scaffold supplies the tools and execution loop around a model. The Kimi model and the agent application running it are separate parts of an evaluated system. ## Check your understanding A K2 benchmark is cited to justify a K3 API deployment. Which differences in model, license, support lifecycle, reasoning settings, and harness would you check before treating that result as evidence for the new workflow? ## Also called Kimi model ## Related terms - [Foundation model](https://darkfactory.dev/glossary/foundation-model) - [Open-weight model](https://darkfactory.dev/glossary/open-weight-model) - [Mixture of experts (MoE)](https://darkfactory.dev/glossary/mixture-of-experts) - [Multimodal model](https://darkfactory.dev/glossary/multimodal-model) - [Reasoning model](https://darkfactory.dev/glossary/reasoning-model) - [Context window](https://darkfactory.dev/glossary/context-window) - [Function calling](https://darkfactory.dev/glossary/function-calling) - [Agent scaffold](https://darkfactory.dev/glossary/scaffold) - [Evaluation (eval)](https://darkfactory.dev/glossary/evaluation) ## Related factory areas - [Model selection, routing & budgets](https://darkfactory.dev/factory/model-routing-budgets) - [Tools & project interfaces](https://darkfactory.dev/factory/tools-interfaces) - [Execution environments, identity & secrets](https://darkfactory.dev/factory/execution-environments) ## Evidence and further reading - [Kimi Team: Kimi k1.5](https://arxiv.org/abs/2501.12599v4) - [Kimi Team: Kimi K2](https://arxiv.org/abs/2507.20534v2) - [Moonshot: Kimi K2 repository](https://github.com/MoonshotAI/Kimi-K2) - [Moonshot: Kimi K2 license](https://github.com/MoonshotAI/Kimi-K2/blob/main/LICENSE) - [Moonshot: Kimi K3 repository](https://github.com/MoonshotAI/Kimi-K3) - [Kimi Team: Kimi K3](https://arxiv.org/abs/2607.24653v2) - [Moonshot: Kimi K3 license](https://github.com/MoonshotAI/Kimi-K3/blob/main/LICENSE) - [Moonshot: Kimi API model list](https://platform.kimi.ai/docs/models) - [Moonshot: Kimi thinking models](https://platform.kimi.ai/docs/guide/use-thinking-models)