Kimi AI Review 2026

What is Kimi AI?

In July 2026 Moonshot AI put a 2.8-trillion-parameter model called K3 behind the Kimi chat box, and the API version of that model accepts 1,048,576 tokens in a single request. That figure, roughly a few thick annual reports at once, is the reason this Kimi AI by Moonshot long context review exists. The questions Indian founders and developers actually ask are narrower: how much of that context reaches the free app, what the $19 plan buys, how the open weights can be used commercially, and where your files end up.

A Beijing lab that has bet on long inputs since 2023

Moonshot AI was founded in March 2023, and the privacy policy on kimi.com names Beijing Moonshot AI Technology Co., Ltd. in Haidian District, Beijing, as the operator of the consumer app. Kimi went into closed beta in October 2023 and opened to the public on 16 November that year with a 128,000-token window, which was large at the time. A March 2024 beta stretched that to about 2 million characters, and the long context window became the product's calling card in China.

The model releases came quickly after that. Kimi k1.5 arrived on 20 January 2025, with Moonshot claiming it matched OpenAI's o1 on maths, coding and multimodal reasoning. Kimi K2 followed in July 2025 as open weights: a mixture of experts design with 1 trillion total parameters and 32 billion active per token. K2 Thinking, K2.5, K2.6 and the coding-focused K2.7-Code came over the next twelve months, before K3 in July 2026. For a Chinese AI company, Moonshot has been unusually open: its GitHub organisation lists 42 repositories, including Kimi-K3 and the kimi-code CLI.

Three models in the picker: K3, K2.6 and K3 Swarm

The Kimi help centre lists three choices in the app today. K3 is billed as the most capable option for chat and agent work, with Low, High and Max thinking levels. K2.6 is the quicker option for plain questions and has Standard and High thinking. K3 Swarm, with the same three levels, is built for wide web search and batch jobs. There is no web search switch: Kimi decides on its own when to look things up.

Around the models sits a set of agentic tools. The General Agent builds websites, slides, spreadsheets and documents, and runs Deep Research. Agent Swarm, per the same help page, coordinates up to 300 sub-agents and more than 4,000 parallel tool calls. Kimi Code covers developers with a CLI and a VS Code extension, Kimi Claw is a cloud automation service with 5,000+ built-in skills, and Kimi Work is a local agent for office tasks. Voice calls, photo problem-solving and reusable prompt presets round out the chat side. Older guides still talk about a separate researcher mode (Kimi-Researcher launched in June 2025) and OK Computer; those jobs now sit under Deep Research and the General Agent.

From Kimi K2 model benchmarks to the K3 table Moonshot published

K2 had its own score sheet in 2025, but the figures that matter now are on the official Kimi K3 model card on Hugging Face. These are Moonshot's claims, run at max reasoning effort with its own test setups, and nobody outside the company has reproduced them for this review.

• Reasoning: 93.5 on GPQA Diamond; 43.5 on HLE-Full without tools and 56.0 with them.
• Coding: 88.3 on Terminal-Bench 2.1, 81.2 on FrontierSWE and 42.0 on SWE-Marathon.
• Agentic work: 91.2 on BrowseComp, 94.5 on MCPMark-Verified and 84.8 on OSWorld-Verified.
• Documents: 91.1 on OmniDocBench and 94.6 on Harvey Lab-AA, a legal task set.

The card compares K3 with Claude Fable 5, GPT-5.6 Sol, Claude Opus 4.8, GPT-5.5 and GLM-5.2, and it does not hide the losses. K3 has the top BrowseComp and MCPMark-Verified scores in that table. It is within a point of GPT-5.6 Sol on GPQA Diamond (94.1) and Terminal-Bench 2.1 (88.8), and well behind Claude Fable 5 on FrontierSWE (86.6). With tools on HLE-Full, Claude Fable 5 (63.0), GPT-5.6 Sol (58.0) and Claude Opus 4.8 (57.9) all finish ahead. Read at face value, the table says K3 is strongest at browsing and tool calls, which suits research agents more than long coding sessions.

Is Kimi AI free to use, and where does the credit pool run dry?

Signing up costs nothing: Kimi runs a free tier with limits Moonshot does not publish, and the help centre's pricing pages list only the paid plans. That matters because every heavy feature on a paid plan draws on one pool. According to Kimi's membership rules, agent tasks, Kimi Code, Kimi Work, Kimi Claw and image generation all deduct from the same monthly credits, and K2.6 used in chat also bills against member credits. When the balance hits zero, a running task finishes but new ones are blocked until the refresh or an upgrade.

Use the free tier to test chat and a document or two. If you find yourself wanting several Deep Research reports or slide decks a week, that is the point to look at a paid plan.

What $19 a month gets you, next to the K3 API rates

Kimi's international help page (pricing screenshot dated 19 October 2026) lists four new plans in USD. No INR price exists, so Indian buyers pay in dollars.

All four plans include Kimi Code with a rolling 5-hour limit and no weekly cap. Credits refresh monthly and do not roll over. Mainland China users see a different ladder in yuan: Go ¥49, Plus ¥99, Pro ¥199, Max ¥699 and Ultra ¥1,399 a month, and older subscribers on Moderato, Allegretto, Allegro or Vivace keep their legacy terms.

On the developer side, the API is pay-as-you-go with taxes added at checkout. K3 writes to its prompt cache are billed separately at $3.00 per million tokens for the default 5-minute lifetime, or $6.00 for one hour. A coding variant, kimi-k2.7-code, has the same $0.95 and $4.00 list rates as K2.6 but charges $0.19 for cache hits. The file extraction and storage endpoints are listed as temporarily free.

Kimi AI long document analysis: 50 files, 100 MB each

This is the job Kimi was built for. According to the help centre, one session accepts up to 50 files of up to 100 MB each, in PDF, Word, Excel, PowerPoint, TXT, image and video formats. A chartered accountant could drop in a year of GST returns plus the auditor's notes and ask for mismatches; a policy team could load a draft bill and three earlier versions and ask what changed in clause 12.

Two cautions. The 1M-token figure is published for the kimi-k3 API model; Moonshot's help pages do not state the exact context each app plan exposes, so very large uploads may still be trimmed. And as a document summarizer it is only as good as the source text: scanned PDFs with poor OCR give poor answers, so ask Kimi to quote the page and line it relied on and spot-check those quotes.

Kimi AI vs ChatGPT, with DeepSeek as the budget third

ChatGPT is the more finished product for teams, with workspace controls, image generation and UPI billing. Kimi is cheaper per dollar of agent work and gives you weights you can run yourself. DeepSeek costs nothing for chat and less on the API, but has no agent suite on Kimi's scale. If your work is mostly reading long PDFs and producing research briefs, Kimi is the one to trial first; for client-facing teams that need admin features, ChatGPT is safer.

The Kimi K3 License: free to build on, until $20 million

Kimi K2 shipped under a modified MIT licence. K3 uses a new Kimi K3 License that adds two conditions. A company running a model-as-a-service business needs a separate agreement with Moonshot before commercial use once its revenue, counted with affiliates, passes US$20 million in any 12 consecutive months. Any commercial product with more than 100 million monthly users or more than US$20 million in monthly revenue must show "Kimi K3" prominently in its interface. Internal use is exempt, as is use through Moonshot's own products or certified inference partners. For nearly every Indian startup this is effectively open source, though self-hosting a 2.8T model in MXFP4 still needs a serious GPU cluster running vLLM or SGLang.

Your files, Beijing servers and the training opt-out

The kimi.com privacy policy (effective 31 August 2026) says personal information is stored in mainland China and that inputs and outputs may be used to improve models after de-identification. There is no settings toggle: you opt out by contacting customer service. The developer platform is different. Its policy (last updated 30 April 2025) names Moonshot AI Pte. Ltd. in Singapore and says data sits on Singapore servers, though content can still be used to train models.

Kimi works in India, on the web, iOS and Android, and we found no ban on it. Government staff should remember the Finance Ministry's February 2025 advice against AI tools on office devices. For agencies, the safe line is simple: no client contracts, KYC data or unreleased financials in the hosted app. Run the open weights on your own servers if that data has to be processed.

Rough edges a paying user will hit

• Plan names have changed twice in a year, and Chinese and international plans differ.
• Free-tier limits and credit costs per task are not published.
• The API now lives at platform.kimi.ai while many tutorials still point to platform.moonshot.ai.
• No India pricing, team admin console or published compliance certifications.

Verdict of this Kimi AI by Moonshot long context review for Indian buyers

Start free: upload the longest PDF you work with and ask for a cited summary, then run one Deep Research task on your market and compare it with what you already know. If both hold up, Plus at $19 is fair value for heavy readers and solo developers who want Kimi Code. Developers should test kimi-k2.6 before kimi-k3, since its output tokens cost about a quarter as much.