Kimi (Moonshot)
Type: app · Vendor: Moonshot AI · Language: Web product + API · License: proprietary (Kimi product) / open-weight MIT (Kimi K2) · Status: active · Status in practice: mature
Moonshot's Kimi assistant and the open-weight Kimi K2 model family — a long-context (256K) MoE model explicitly trained for tool calling and agentic behaviour, surfaced through the Kimi consumer product line (Deep Research, Agent Swarm, Kimi Code, OK Computer).
Description. Kimi began in 2023 as Moonshot AI's chatbot known for 128K-token context; subsequent versions raised the window to 256K and added multimodal text/image/video input. Kimi K2 (released July 2025) is the open-weight 1T-total / 32B-active MoE model behind the product, explicitly trained for tool use and autonomous problem-solving and licensed permissively. The Kimi consumer surface bundles long-context chat with Deep Research, Agent Swarm, Slides/Docs/Sheets/Websites generation, and a sub-product called OK Computer that can build multi-page sites and editable slides from prompts and process up to 1M rows of input data. Public English documentation of the underlying API's tool-call schema redirects between several domains (platform.moonshot.cn → platform.kimi.com → platform.kimi.ai) and several pages were not directly fetchable; primary technical evidence here comes from the Kimi K2 model documentation.
Agent loop shape. Two surfaces. (1) Kimi consumer app: Moonshot-hosted assistant; user starts a chat or invokes a sub-product (Deep Research, Agent Swarm, OK Computer); model autonomously calls internal tools and produces multi-modal artifacts. (2) Kimi K2 API: standard chat-completions tool-call loop where the model autonomously decides when and how to invoke tools, designed for drop-in general chat and agentic experiences.
Primary use cases
- long-context document analysis and Q&A
- agentic tool-calling deployments via Kimi K2 open weights
- Deep Research and Agent Swarm consumer workflows
- AI-built slides, websites, sheets and docs from prompts
Key concepts
- Kimi K2 → tool-use (docs) — Open-weight 1T total / 32B active MoE model (384 experts, 8 routed per token) trained on 15.5T tokens; designed for tool use and autonomous problem-solving.
- Long-context window (256K) → context-window-packing (docs) — Kimi K2 / Kimi 2.6 supports 256K tokens of context, with K2 documentation also listing 128K base.
- Agent Swarm (docs) — Product feature in the Kimi consumer app advertised on the home page alongside Deep Research, Kimi Code and others.
- Deep Research → agentic-rag (docs) — Kimi consumer research mode; the older Kimi Explore Edition reportedly deeply reads 500+ pages per search.
- OK Computer (docs) — Kimi sub-product that builds multi-page websites and editable slides from prompts and processes up to 1M rows of input data.
Patterns this app implements —
- ★★Agentic RAG
Kimi Explore / Deep Research is positioned as autonomous AI-powered search. K2.5 is documented as equipped with search, code-interpreter and web-browsing tools and explicitly enables deep research an…
- ★Browser Agent
Kimi K2.5 ships Kimi WebBridge — a browser extension for AI agents — and turns prompts into full websites. K2.5 is positioned as a visual agentic model with web-browsing tools.
- ★★Context Window Packing
Long context is Kimi's signature: initial 128K, later 256K.
- ★★Tool Use
Kimi K2 is explicitly designed for tool use; the model autonomously decides when and how to invoke tools.
References
Provenance
- Last analyzed:
- Last updated:
- Verification status: verified