Kimi

Kimi (kimi.com) is an AI assistant operated by Beijing Moonshot AI: type a question or task into a single input box and it handles Q&A, web search, deep research, slides/document/spreadsheet/website generation, and coding. It launched in October 2023 as a chat product known for supporting 200,000 Chinese characters of context; today the official positioning has widened to a full-stack assistant covering chat, agent tasks, coding, and local desktop collaboration. Two verifiable differences from most peers: first, the underlying models are self-developed and open-weight — Kimi K2 (1 trillion parameters, July 2025) shipped under a modified MIT license, and Kimi K3, released on July 16, 2026, is at 2.8 trillion parameters the largest open-weight model to date; second, the product surface is complete — one account spans web/app chat, the Kimi Code coding assistant, the Kimi Work desktop agent, the Kimi Claw cloud agent, an API platform, and an enterprise tier.

At a glance

URL: https://www.kimi.com (mainland China); https://www.kimi.ai (international); the legacy domain kimi.moonshot.cn now redirects to kimi.com (tested 2026-09-17)
Type: AI assistant (chat + agent tasks) + developer API platform + enterprise tier
Cost: free to use after sign-up, with a basic quota; four personal membership tiers at ¥49/99/199/699 per month (annual billing works out to ¥39/79/159/559 per month); as of 2026-09-17 the purchase buttons read "reserve subscription" — see the official pricing page and help center; the API is billed per token
Account required: chat requires sign-in (we tested: sending a message while logged out pops up a login dialog); kimi.com accepts WeChat QR scan or mainland-China phone SMS, kimi.ai accepts a phone number or a Google account; users must be 18 or older, or have guardian consent (Terms of Service, effective 2026-08-31)
Interface language: Simplified Chinese by default; settings offer 13 languages including English, Traditional Chinese, Korean, French, German, Russian, and Spanish — Japanese is not among them (official FAQ); the international site kimi.ai additionally has Japanese and Arabic page versions
Launched: October 2023 (Kimi Chat's first release; opened to the public on November 16, 2023)

Background

Moonshot AI was founded in Beijing in March 2023 by three Tsinghua alumni — Yang Zhilin, Zhou Xinyu, and Wu Yuxin. The company name marks the 50th anniversary of Pink Floyd's The Dark Side of the Moon; it is headquartered in the JD Technology Building in Haidian, Beijing (Wikipedia, official About page). Founder Yang Zhilin previously worked at Google Brain and Meta and contributed to Huawei's Pangu LLM; "Kimi" comes from his English name.

The product's early label was long context: the first Kimi Chat in October 2023 supported roughly 200,000 Chinese characters, and in March 2024 the company announced a 2-million-character context beta — that same month the service went down for two days under surging traffic and issued a public apology (Wikipedia, citing Shanghai Securities News and the FT). The product then moved toward agents: an auto-searching "Explorer" version in October 2024, the Kimi-Researcher deep-research agent in June 2025, and the OK Computer agent mode (which generates websites and slide decks) in September 2025.

Model milestones, all per official release pages: July 2025 — Kimi K2 open-sourced (1T total / 32B active MoE, trained on 15.5T tokens); November 2025 — K2 Thinking; January 2026 — K2.5 (native vision, introducing agent swarms and Kimi Code); April 2026 — K2.6 open-sourced, focused on long-horizon coding; June 2026 — K2.7 Code; July 16, 2026 — K3. Separately, as of 2026-09-17 the logged-out homepage of the international site kimi.ai already shows a "K2.8" option in the model picker (observed in the UI), but no official announcement exists.

Funding has clearly accelerated over the past two years: February 2024 — roughly $1B led by Alibaba at a$ 2.5B valuation; May 2026 — a ~ $2B round at a valuation above$ 20B led by Meituan's strategic investment arm, with annualized revenue of about $200M disclosed for April 2026 ([TechCrunch](https://techcrunch.com/2026/05/07/chinas-moonshot-ai-raises-2b-at-20b-valuation-as-demand-for-open-source-ai-skyrockets/)); late July 2026 — another$ 3.5B at a $35B valuation, led by China's state AI industry investment fund (Bloomberg, via Silicon Republic and others), alongside reports of a planned Hong Kong IPO.

Core capabilities: from chat to deliverables

Kimi's logged-out homepage: a chat input box with shortcuts for slides, swarms, deep research, docs, websites, sheets, and design

Kimi's homepage is the workbench (above): model and thinking-effort selectors sit above the input box, and the shortcuts below map to seven agent capabilities. Per the official help center (as of 2026-09-17):

Three model tiers: K2.6 for fast Q&A with a ~128K-token single-turn context (roughly 50–60k Chinese characters), text output only; K3, the most capable, can produce editable .pptx/.docx/.xlsx/.pdf end to end, and its 1M-token ultra-long conversations are reserved for the top membership tier; K3 Swarm targets massive search and batch tasks — officially up to 300 sub-agents and more than 4,000 parallel tool calls.
Thinking effort: Standard / Advanced / Extreme — higher tiers reason more and consume more tokens; web search is decided by the model itself, with no manual toggle.
Agent feature matrix (features page): Deep Research (long cited reports with charts), Docs (Word/PDF generation and review), Sheets (Excel with formulas and pivot tables), Slides, Websites (database-backed and deployable), and Design; a plugin marketplace adds image, video, and audio generation.
File handling: PDF, Word, Excel, PPT, images, TXT, and video, up to 100MB per file and 50 files per session (capability FAQ).

Kimi's features page: cards for the five agent capabilities — deep research, docs, sheets, slides, and websites

One account also unlocks several separate product lines (official onboarding docs): Kimi Code, a developer-facing coding assistant (CLI + VS Code extension, also usable inside third-party tools like Claude Code and Roo Code via API key); Kimi Work, a local agent inside the Mac/Windows desktop client that can operate your desktop with permission (Kimi Computer Use), with widgets and kanban boards; Kimi Claw, a cloud-hosted OpenClaw agent with a built-in library of 5,000+ skills and Feishu integration; plus iOS, Android, and HarmonyOS apps and a browser extension.

The Kimi K3 model page: 2.8T parameters, 1M-token context, native vision, and architecture notes

K3 is the current flagship (above): a 2.8T-parameter MoE with 896 experts and 16 active per token, built on the in-house Kimi Delta Attention and Attention Residuals, with native image, video, and document understanding and a 1M-token context window. The official blog claims it trails only Claude Fable 5 and GPT-5.6 Sol on coding, reasoning, and knowledge-work benchmarks, and shows two long-horizon demos: autonomously completing a 48-hour chip-design flow for the company's own nano model, and cutting its own trailer from 56 clips.

Accounts and pricing

As of 2026-09-17, the official purchase page and help center list four personal tiers (all named after musical tempos; prices in CNY):

Tier	Monthly	Annual (effective monthly)	Positioning
Andante	¥49/mo	¥39/mo (¥468/yr)	Everyday use
Moderato	¥99/mo	¥79/mo (¥948/yr)	Productivity
Allegretto	¥199/mo	¥159/mo (¥1,908/yr)	Professional
Allegro	¥699/mo	¥559/mo (¥6,708/yr)	All-inclusive

Kimi's membership purchase page: Andante ¥49, Moderato ¥99, Allegretto ¥199, and Allegro ¥699 per month; buttons read "reserve subscription"

The billing model deserves attention: all membership features (agent tasks, deep research, slides, Kimi Code, Kimi Work, Kimi Claw, and so on) draw from one shared quota pool, deducted by actual token consumption — even K2.6 chat counts; free users get a small quota too. Official reference points: a simple slide deck costs about 1–2% of a monthly quota, and one deep-research run about 5–10% (quota rules). Quotas refresh monthly and do not roll over; Kimi Code has an additional five-hour weekly rate window. Higher tiers mainly multiply the quota (2x/4x/10x), agent parallelism, and swarm runs, and unlock extras such as goal mode, a dedicated Kimi Claw, and million-token conversations (Allegro only). When the quota runs out, a pay-as-you-go "fuel pack" is available (minimum top-up ¥25, balance never expires).

Two time-sensitive notes. First, as of 2026-09-17 all four tiers show "reserve subscription" buttons, a banner says a new membership plan is coming, and the Kimi Code benefits page explains the new system will split Kimi membership and Kimi Code benefits into separate purchases (existing subscribers are unaffected) — the pricing above reflects that day's pages and may change. Second, for teams there is Kimi Business: ¥4,200 per seat per year, two seats minimum, with a default commitment not to train on enterprise data (enterprise intro).

For developers, the Kimi API platform (platform.kimi.com / platform.kimi.ai) bills per token; official K3 pricing is $3.00 per million input tokens ($ 0.30 on cache hits) and $15.00 per million output tokens (K3 blog). The API is compatible with both OpenAI and Anthropic formats. New users registering with a mainland-China phone number and completing real-name verification receive a ¥15 voucher (valid for 3 months, not usable for K3) — see the free-trial notes.

Open-weight models and licensing

K-series weights are published on Hugging Face (moonshotai), under two license families:

K2 series (K2, K2-0905, K2 Thinking): a modified MIT license — one added clause requiring prominent "Kimi K2" attribution once a product exceeds 100 million MAU or $20M in monthly revenue (LICENSE text).
K3: the custom Kimi K3 License (we read the full text) — free use, modification, and deployment, but providers running "model-as-a-service" businesses with annual revenue above $20M must sign a separate agreement with Moonshot before commercial use; the same 100M MAU /$ 20M monthly revenue attribution threshold applies; internal use and use through official channels are exempt.

For scale: as of 2026-09-17, K3 has roughly 2.236 million downloads and 11k likes on Hugging Face. Intermediate versions (K2.5, K2.6, K2.7 Code) are also open-weight, and research projects including Kimi-VL, Kimi-Linear, and the Muon optimizer are on GitHub (MoonshotAI).

Terms and content policy highlights

Per the Terms of Service effective 2026-08-31: users keep the rights to their inputs and outputs; the platform receives a free license to use inputs, outputs, and feedback for model optimization, with an opt-out via customer support; generated content carries explicit and implicit AI labels by default under China's AI-generated-content labeling rules, and the labels may not be removed; outputs are for reference only and do not constitute medical, legal, or financial advice; disputes are governed by Chinese law in the courts of Beijing Moonshot's domicile.

Where it fits

Long-document work: upload contracts, papers, or financial reports (100MB per file, 50 at a time) for summaries, Q&A, and review — Kimi's original strength.
Office deliverables: one-sentence generation of editable slides, Excel, Word, or database-backed websites — good for fast first drafts.
Deep research: cited industry research or literature reviews via Deep Research, or batch retrieval with K3 Swarm.
Developers: a Kimi Code subscription covers the CLI and IDE extension, or call the API directly; teams needing private deployment can download K2/K3 weights and self-host.
Enterprises: Kimi Business adds member management and the no-training-by-default commitment.

Limitations

Everything is metered: membership is a shared token quota pool, not "unlimited" — heavy agent tasks (deep research, swarms) burn fast, and one deep-research run can take 5–10% of a monthly quota; the free quota only covers light trials.
Flagship features are locked to the top tier: K3's 1M-token conversations require the ¥699/month Allegro plan; swarms and goal mode also have tier gates.
Pricing is mid-transition: as of 2026-09-17 new subscriptions are in "reserve" mode, a new plan has been announced that splits Kimi and Kimi Code benefits, and the prices here reflect that day only.
Compliance and content boundaries: as a Chinese service, generated content is subject to Chinese regulation with mandatory AI labeling, and the terms list detailed prohibited-content categories.
The distillation dispute: in February 2026 Anthropic publicly accused Moonshot and others of distilling Claude via fraudulent accounts; on September 10, 2026 Anthropic went further with a report claiming Moonshot had routed some Kimi user requests to Claude and returned its answers (nearly 300,000 requests over 10 days), including sensitive user data. A Moonshot employee previously rebutted the distillation claims by noting K3 shipped only 15 days after Fable; as of 2026-09-17 we found no official response to the September report. These remain one-sided allegations without independent verification, but the data-security claims are worth enterprise users' attention.
Output reliability: the terms explicitly disclaim accuracy; anything important needs human review.

Alternatives

DeepSeek (listed here as deepseek): also a Chinese open-weight model plus assistant, but a different route — DeepSeek's web/app chat is free, while Kimi is "free base + membership".
ChatGPT (listed here as chatgpt): OpenAI's counterpart, with a more complete ecosystem and plugin system, but no open weights.
Doubao: ByteDance's AI assistant, the consumer-scale leader in mainland China, aimed at everyday scenarios.