Kimi

Overview

Kimi is the conversational AI assistant from Moonshot AI, a Beijing-based startup founded in March 2023 by Yang Zhilin and two Tsinghua University classmates. Launched in late 2023, Kimi built its reputation on long-context understanding: its first version handled a lossless context of 128,000 tokens, and later K2 releases reached a 256K-token window. It runs on Moonshot's own open-weight K2 model family rather than a third-party model.

What it does

The base assistant answers questions, analyzes long documents and uploaded files, writes and edits text and code, and searches the live web, all in a chat interface available on the web, app, and via API. On top of that, Moonshot has layered agentic surfaces: Kimi Researcher plans and runs multi-step web research into a report; OK Computer builds slides, websites, sheets, and docs from a prompt; Kimi Code generates and edits code across languages; and Agent Swarm coordinates multiple sub-agents in a shared workspace with a coordinator managing tasks and dependencies. These surfaces carry out multi-step work under the user's review.

Autonomy note

Treat Kimi itself as an assistant: it produces output when asked rather than acting on its own. The Researcher, OK Computer, Kimi Code, and Agent Swarm surfaces do multi-step work, but under the user's oversight, which makes them supervised rather than fully autonomous.

Integrations & setup

Used via the web at kimi.com, mobile apps, and the Moonshot/Kimi API platform. Because the K2 series is open-sourced under a modified MIT license, the models can also be self-hosted from Hugging Face. The API supports function calling and standard REST access.

Pricing

Freemium. There is a free tier with rate limits, plus paid subscription tiers named after musical tempo markings: Moderato (reported around $19/month), Allegretto, Allegro, and Vivace, with higher tiers unlocking more agent credits, more Kimi Code usage, and larger Agent Swarm parallelism. API access is billed per token (reported around $0.55 per million input tokens and $2.65 per million output tokens). Exact prices change frequently; check the official pricing page (third-party figures are not official).

Best for / not for

Best for users who need long-document analysis, multi-step research, or open-weight models they can self-host, and for developers who want competitively priced API access to a frontier open model. Less suited to buyers who need a hands-off autonomous worker out of the box, or to organizations with data-residency concerns about a China-based vendor.

Alternatives

ChatGPT, Claude, and Perplexity are the main Western general-assistant and research rivals; DeepSeek and Qwen are the closest open-weight, China-based alternatives.