AIWiki
Malaysia

Kimi

A conversational AI assistant and long-context large language model developed by Moonshot AI, a Beijing startup, known for its industry-leading context window lengths and strong performance on agentic reasoning tasks.

4 min readLast updated June 2026Models

Kimi is a conversational AI assistant and family of large language models developed by Moonshot AI (月之暗面, meaning "dark side of the moon"), a Beijing-based AI startup founded in March 2023. Moonshot AI was established by Yang Zhilin, a former researcher at Carnegie Mellon University and Tsinghua University, along with co-founders from leading Chinese and international AI institutions.

Kimi is primarily distinguished by its emphasis on long-context processing — the ability to reason coherently over very long documents or conversations — which has been a core research focus for the Moonshot team from its inception.

Company and Funding

Moonshot AI raised a USD 1 billion Series B funding round in February 2024, one of the largest single funding rounds for a Chinese AI startup, with investors including Alibaba, HongShan (Sequoia China), and Zhen Fund. This valuation placed Moonshot AI among China's most highly valued AI companies. By early 2025 the company's valuation was reported at approximately USD 3 billion.

Kimi K1 and Early Long-Context Work

The initial Kimi product launched in late 2023 with a 128,000-token context window — an industry-leading specification at the time for a publicly accessible consumer AI product. This capability, enabling users to upload and query entire books, lengthy research papers, or extensive codebases within a single session, became Kimi's primary market differentiator and drove rapid adoption among researchers, analysts, and developers in China.

Kimi K2 Series

The Kimi K2 family, released in mid-2025, represented a significant architectural advance. Kimi K2 features a 1-trillion-parameter Mixture-of-Experts architecture with 32 billion active parameters, combining long-context capability with strong agentic reasoning performance. The K2 series emphasises practical task completion rather than benchmark performance alone, positioning Kimi as an AI agent capable of autonomously executing multi-step tasks such as coding projects, research synthesis, and document workflows.

Kimi K2 Instruct, released as an open-weight model on Hugging Face, supports 256,000-token contexts and demonstrates competitive performance on coding, mathematics, and instruction-following benchmarks. The open release reflects Moonshot AI's strategy of building a developer ecosystem around the Kimi model family.

Kimi K2 Thinking (November 2025) added explicit chain-of-thought reasoning to the K2 base, improving performance on complex logical, mathematical, and scientific problems through extended inference-time computation.

Kimi K2.5 (January 2026) extended the series to multimodal capabilities, incorporating a 400-million-parameter vision encoder named MoonViT that enables native image understanding alongside the existing long-context text capabilities. K2.5 supports 256K context with tool calling.

Kimi K2.6 refined long-context coding stability, addressing failure modes in extended software engineering tasks.

Consumer Product

The Kimi consumer application is available via web browser and mobile app, offering features including document upload and Q&A (accommodating files up to several hundred pages), extended-session conversation, web search integration, and code assistance. Kimi's interface is available primarily in Chinese and English.

See Also

References

References

  1. Wikipedia. (2025). Kimi (chatbot). en.wikipedia.org.
  2. Wikipedia. (2025). Moonshot AI. en.wikipedia.org.
  3. Moonshot AI. (2025). Kimi K2: Open agentic intelligence. moonshotai.github.io/Kimi-K2.
  4. Hugging Face. (2026). moonshotai/Kimi-K2-Instruct model card. huggingface.co.
  5. Medium. (2025). Kimi K2 thinking model: Open agentic LLM that redefines intelligence. medium.com/data-and-beyond.