Moonshot AI: Kimi Models, Products and License Terms

Moonshot AI develops Kimi models and applications. Explore its public checkpoints, hosted services and release-specific licensing conditions.

kindCompany
founded2023

Latest

Updated

Moonshot is the company, Kimi the product family

Moonshot AI’s company page places its founding in spring 2023 and gives a Beijing address. Kimi names its model and application work, with separate consumer, business and API offerings. Those access paths should not be treated as separate companies or as identical ways to obtain a model.

Public checkpoints for agent-oriented models

The Kimi K2 release included base and instruction-tuned checkpoints. The model uses a mixture of experts, with one trillion total parameters and 32 billion active parameters. Its model card describes tool-use behavior and deployment interfaces, while the host application still has to execute tool calls.

A precise license matters more than a family label

K2’s weights use a modified MIT license. The added clause requires prominent attribution for commercial products or services above specified monthly user or revenue thresholds. Other Kimi versions and hosted services need their own terms checked. Public weights support independent deployment and inspection, without making the developer’s benchmark tables a guarantee for every application.

Sources

  1. [1]
    Moonshot AI

    Moonshot AI · retrieved Oct 23, 2026

  2. [2]
    Moonshot AI

    Moonshot AI · retrieved Oct 23, 2026

  3. [3]
    moonshotai/Kimi-K2-Instruct

    Hugging Face · Jul 11, 2025

  4. [4]
    Modified MIT License

    Moonshot AI · retrieved Oct 23, 2026

Moonshot AI history

2026

│
Moonshot launches Kimi K3.1 Pro

Moonshot AI released Kimi K3.1 Pro and its faster Highspeed variant through its apps, coding tools and API. Moonshot says the 3T-class model supports a million-token context; downloadable weights were promised by November 4.

Moonshot AI
Model
│
Anthropic accuses seven Chinese labs of distilling Claude

With Anthropic, Alibaba, DeepSeek, Xiaomi

Anthropic's September misuse report named Alibaba, Moonshot, DeepSeek, Zhipu, Xiaomi, SenseTime and MiniMax, and said Moonshot, DeepSeek and Xiaomi fed their own users' conversations into Claude. It also described a blocked request for help with gain-of-function work on chikungunya virus.

Anthropic
Research
■
Moonshot publishes Kimi K3's weights

2.8 trillion parameters with 104 billion active, the largest open-weight model anyone had released. Downloading it means moving about 1.4 terabytes, which is a moat of a sort.

Hugging Face
Open source
■
Moonshot ships Kimi K2.5 with a swarm mode

A trillion-parameter mixture of experts trained on 15 trillion mixed visual and text tokens, released open weight, with a mode that lets it direct up to a hundred sub-agents at once. Moonshot says it takes the top score on HLE-Full and beats GPT-5.2 on several benchmarks.

SiliconANGLE
Model

2025

■
Kimi K2 Thinking tops the open-weight field

A trillion-parameter reasoning model from Moonshot AI, released with open weights and beating several closed frontier models on agentic benchmarks. The US government's AI standards center published its own evaluation five weeks later.

Moonshot AI
Model
│
Moonshot AI publishes the Kimi Linear attention architecture

Kimi Linear replaces most attention layers with a hybrid linear mechanism called Kimi Delta Attention, cutting KV cache use by up to 75 percent and lifting decoding throughput up to 6 times at a million token context while matching full attention on benchmarks.

arXiv
Research
│
Moonshot AI ships Kimi k1.5, an o1-class reasoning model

Moonshot AI released Kimi k1.5, trained with reinforcement learning over context windows up to 128k tokens, and reported math and coding scores matching OpenAI's o1 without Monte Carlo tree search or a separate reward model. Code and a technical report shipped alongside it.

arXiv
Model