Moonshot AI

Kimi

Provider

Moonshot AI

Price

$0-199/mo

Platforms

Web, iOS, Android

Launched

Oct 2023

Updated November 2026

Kimi is the AI assistant from Chinese lab Moonshot AI, and its signature strength is long-document analysis and agentic research. Since 23 October 2026 the app has run Kimi K3.1 Pro, a 2.8-trillion-parameter model that Artificial Analysis scored 55 on its Intelligence Index (v4.3.2) at launch (Moonshot). It followed Kimi K3, which Artificial Analysis ranked 4th of all models at its July 2026 launch, ahead of Claude Opus 4.8, then the strongest result any Chinese lab had posted. Kimi is free to use for chat, web search and long-document work, with paid plans from $19/month unlocking the flagship model and the heaviest agentic features. The single biggest caveat is jurisdiction: Kimi is a China-based service that stores data on servers in China, so for sensitive or regulated work you should self-host the open-weight Kimi models or use a Western inference provider instead.

What sets Kimi apart from a normal chatbot is its agent mode, called OK Computer, which gets its own virtual computer to research, build websites, generate editable slides and process up to a million rows of data in a single run. See our rankings of the best AI apps and the best AI models.

Quick stats

Pricing breakdown

Kimi’s apps are free to use with general rate limits, and the paid plans are named after musical tempo markings. All prices below are in USD; Moonshot also prices in Chinese yuan and figures vary by region and purchase channel, so confirm the current rate on the official pricing page before subscribing.

Free — $0

The free tier is genuinely useful and is the main reason Kimi is worth trying. You get:

• Free chat, web search and file uploads — drop in a long PDF and work through it
• Long-document analysis — upload many files at once, up to 10MB each
• Kimi-Researcher and OK Computer — the research agent and agent mode, with usage limits
• A standard Kimi model — competent for everyday tasks

Limitations: The flagship model and the largest agent and research quotas are reserved for paid plans; free uploads are capped at 10MB per file.

Worth it for: Students, researchers and anyone wanting to test Kimi’s long-document and research strengths before paying.

Moderato — $19/month

The entry subscription, and the cheapest way to get the flagship. Everything in Free, plus:

• Kimi K3 in the chat interface — with a 256K context, as Moonshot listed the plan in July 2026; Moonshot has not said which plans include Kimi K3.1 Pro
• Higher daily usage and agent credits
• Deep Research, Kimi Slides and Kimi Websites
• Kimi Code access — the agentic coding CLI
• 100MB file uploads

Worth it for: Regular users who want the best Kimi model and the agent tools. At $19/month it undercuts the $20 charged by ChatGPT Plus, Claude Pro and Gemini Advanced.

Allegretto — $39/month

For heavier users. Everything in Moderato, plus the full 1M-token context window (for Kimi K3, as listed in July 2026), larger quotas, more Kimi Code and agent credits, and extended OK Computer and Agent Swarm usage.

Worth it for: Power users who lean on Kimi’s agent mode and coding tools through the day.

Allegro — $99/month

A heavier tier between Allegretto and the top plan. Everything in Allegretto, plus much larger usage quotas, heavier Kimi Code and professional-data allowances, and Kimi Claw cloud deployment for running agents in the cloud.

Worth it for: Professionals running sustained agentic and coding workloads.

Vivace — $199/month

The top tier. Everything in Allegro, plus the full Agent Swarm (up to 300 parallel sub-agents) and the largest professional-data and Kimi Code quotas.

Worth it for: Teams and professionals running the largest autonomous workloads.

Pricing varies by region and channel. Reported figures differ slightly by source and purchase channel (for example, Moderato has been listed as low as $15/month), and in-app store prices can differ from web prices. In mainland China, Kimi sells shorter passes — six tiers ranging from about 5.2 yuan for four days to 399 yuan for a year. Always check the live pricing page for your region.

Developer and enterprise access

There is no consumer team plan. Businesses and developers use the Moonshot Open Platform — a pay-as-you-go API where Kimi K3.1 Pro costs $2 per million input tokens and $10 output, Kimi K3 $3 in, $15 out and $0.30 for cached input, and the older K2.6 roughly $0.95 in and $4 out. At its July launch K3’s rates were roughly triple K2.6’s. Because the Kimi models are open-weight — the K2 line under a modified MIT licence, K3 since 26 July 2026 under the Kimi K3 License and K3.1 Pro since 4 November 2026 under the MIT-based Kimi K3.1 Pro License — you can also self-host them or run them through third-party inference providers, the standard route for organisations that need data to stay outside China (K3 and K3.1 Pro are 2.8T parameters, about 1.56 TB, so they need serious hardware).

Models powering Kimi

Kimi was built around long context from the start, and its models are Mixture-of-Experts systems. In the app you chat with the latest model by default; each has its own benchmarks and pricing on the model page.

Kimi K3.1 Pro (October 2026)

Kimi K3.1 Pro has been Moonshot’s flagship since 23 October 2026, in the Kimi apps, Kimi Work, Kimi Code and on the API at $2/$10 per million tokens (Moonshot):

• 2.8 trillion parameters, 104 billion active, with image input and a 1M-token context
• 55 on the Artificial Analysis Intelligence Index (v4.3.2) at launch, at $2.11 per completed index task
• Open weights since 4 November 2026 on Hugging Face, about 1.56 TB in MXFP4, under the MIT-based Kimi K3.1 Pro License, which requires a separate Moonshot agreement for model-as-a-service businesses above $20 million in revenue
• No SWE-bench figure published by Moonshot

See the Kimi K3.1 Pro page for full benchmarks.

Kimi K3 (July 2026)

The previous flagship, released on 16 July 2026:

• 2.8-trillion-parameter MoE, 16 of 896 experts active — nearly three times the size of K2.6
• 1M-token context window — quadruple K2.6’s, though the app caps Moderato at 256K
• Native vision — reads images and video
• Always-on thinking — a reasoning model first; only the max effort level at launch
• 4th overall on independent testing at launch — 57.1 on the Artificial Analysis Intelligence Index v4.1 in July 2026, above Claude Opus 4.8 (55.7) and behind only Claude Fable 5 and GPT-5.6 Sol at the time

Two caveats: Artificial Analysis also measured a higher hallucination rate than its predecessor (51%), and — unusually for a Kimi flagship — the open weights lagged the launch, finally shipping on 26 July 2026 under the Kimi K3 License. See the Kimi K3 page for full benchmarks.

Available to: Moderato and above (256K context); Allegretto and above get the full 1M window, as listed in July 2026.

Kimi K2.6 (April 2026)

The previous flagship, released on 20 April 2026:

• One-trillion-parameter MoE, 32 billion active — large capacity, efficient inference
• 256K-token context window — entire codebases, contracts and document sets in one pass
• Native multimodal — reads images and video through Moonshot’s MoonViT vision encoder
• Agent Swarm — coordinates up to 300 specialised sub-agents across as many as 4,000 steps in a single autonomous run
• Open weights — modified MIT licence, free to download and self-host

On independent testing by Artificial Analysis, K2.6 debuted at #4 across all models on the Intelligence Index v4.0 with a score of 54, three points behind the frontier trio of Claude Opus 4.7, Gemini 3.1 Pro and GPT-5.4 (tied at 57 at the time), and #1 among open-weight models — ahead of the next-best open model, GLM-5.1. Moonshot’s own figures put K2.6 at 58.6% on the memorisation-resistant SWE-bench Pro (which, on that specific benchmark, leads GPT-5.4 at 57.7% and Gemini 3.1 Pro at 54.2%) and roughly 80.2% on SWE-bench Verified. These coding figures are vendor-reported and should be read as a ceiling; the independent Intelligence Index is the better like-for-like measure, and at K2.6’s April launch it trailed the very top US models but led every open model.

Available to: Moderato, Allegretto and Vivace subscribers (the paid plans).

Kimi K2.5 (January 2026)

The first multimodal Kimi, released 27 January 2026. K2.5 added native vision (image and video understanding) via a 400-million-parameter MoonViT encoder, plus instant and thinking modes, and could replicate a website’s user journey from a video demonstration alone.

Kimi K2 Thinking (November 2025)

An open-source reasoning model that can execute 200–300 sequential tool calls autonomously, with a 256K context and native INT4 quantisation for efficiency. It was trained for a reported $4.6 million and was widely described as the leading open-source model at launch.

Earlier models

Kimi K2 (July 2025) was the original open-weight one-trillion-parameter release that drew comparisons to “another DeepSeek moment”. The September 2025 update (K2-Instruct-0905) doubled the context window from 128K to 256K. Before that, Kimi pioneered long context: its 2023 launch supported 128K tokens — the first model to do so — and a 2024 beta handled two million Chinese characters in a single prompt.

Features deep dive

OK Computer — the agent mode

OK Computer, added in September 2025 and named after the Radiohead album, is Kimi’s headline agentic feature. In this mode Kimi gets its own virtual computer and works continuously through research, planning, analysis, design and development:

• Builds multi-page websites and editable slides from a single prompt
• Processes up to one million rows of input data at once
• Outputs text, audio, images and video, not just chat replies

It turns Kimi from a question-and-answer chatbot into a tool that returns finished deliverables.

Agent Swarm

Introduced with K2.6, Agent Swarm scales OK Computer to up to 300 domain-specialised sub-agents executing as many as 4,000 coordinated steps in a single run — up from 100 sub-agents and 1,500 steps in K2.5. It is aimed at large, long-horizon tasks such as building a full-stack application end to end.

Kimi-Researcher and Deep Research

Kimi-Researcher, released June 2025, is an autonomous research agent that goes beyond summarising to systematic analysis, source tracking and evidence mapping. The Deep Research mode plans a query, reads across many sources and returns a structured, cited report — Kimi’s autonomous answer to the deep-research features in ChatGPT and Gemini.

Long-document analysis

Long context is Kimi’s founding strength. The 256K-token window lets you upload dozens of PDFs, documents or slides at once and ask for a summary, comparison or detailed outline, with the model tracking key points across files so follow-up questions stay grounded in the source material. Moonshot’s Explore Edition (October 2024) could “deeply read 500+ pages” in a single search.

Visual coding

K2.6’s MoonViT vision encoder accepts design mockups and screenshots and converts them into code. Describe a website or application in natural language, or hand Kimi a mockup, and it generates a complete full-stack project — frontend, backend, database and authentication.

Kimi Slides, Websites and Office

The app includes built-in tools to turn a prompt or a set of documents into editable slide decks, web pages and office files (PowerPoint, Excel, Word). A newer “document-to-skill” feature lets you turn a reference document into a reusable skill the agent can apply to later tasks.

Web search

Kimi’s web search reads current information and cites its sources, and is available on the free tier. It triggers automatically when a query needs up-to-date facts.

Kimi Code

Kimi Code is Moonshot’s agentic coding tool — a command-line interface that has offered Kimi K3.1 Pro in its model selector since 23 October 2026 and competes with the likes of Claude Code at a lower token cost. It is included from the Moderato plan up and is a major reason Kimi has gained traction with developers. See the best AI for coding for how it stacks up.

Kimi Work — the desktop agent

Kimi Work is Moonshot’s desktop AI-agent product for knowledge workers, bringing OK Computer-style autonomy to a dedicated desktop app rather than the browser. It targets the same goal-in, deliverable-out workflows as rival desktop agents.

Strengths

Long-document and research depth — Kimi was built around long context and remains one of the best tools for reading across large document sets, with a 256K window and an autonomous research agent.

Genuine agentic work — OK Computer and Agent Swarm complete multi-step tasks, build websites and apps, and generate slides and reports, going well beyond chat.

Capability — Artificial Analysis scored Kimi K3.1 Pro 55 on its Intelligence Index (v4.3.2) at its October 2026 launch, and ranked Kimi K3 4th of all models at its July 2026 launch, ahead of Claude Opus 4.8; both models’ weights can be downloaded and self-hosted.

Low cost — The Moderato plan is $19/month, under the $20 of Western rivals, and the API is roughly 80% cheaper than comparable frontier models. The open weights mean zero licensing cost if you self-host.

Open weights — K2.6 (modified MIT), K3 and K3.1 Pro can be downloaded, audited, fine-tuned and self-hosted — the escape hatch from the China-hosting concern.

Strong coding — Kimi leads several open-weight coding benchmarks and Kimi Code is a credible, cheaper alternative to closed agentic coding tools.

Limitations

China data residency — The Kimi app processes and stores data on servers in China, which makes it the wrong choice for sensitive, confidential or regulated work. China’s National Intelligence Law is the underlying concern, the same one that applies to DeepSeek and Qwen.

Flagship is paid — The free tier runs a lighter model; consistent access to the flagship model and the largest agent quotas requires a subscription.

Narrow outside its strengths — Kimi is excellent at long-context and agentic work but less of an all-rounder than ChatGPT or Gemini for everyday multimedia, voice and ecosystem breadth.

Mainly a China-market product — Moonshot has focused on the domestic market and stated it had no plans for dedicated overseas products; kimi.com is internationally accessible, but the app is not localised or marketed the way Western assistants are.

Distillation allegation — In February 2026, Anthropic accused Moonshot (along with DeepSeek and MiniMax) of using fraudulent accounts to harvest Claude data for model training. The claim is an allegation that Moonshot has not accepted, and it has not been independently proven.

Declining consumer reach — Kimi’s monthly active users have fallen from a 2024 peak; by early 2026 it sat well behind the Chinese market leaders (see below).

Vendor benchmarks — Moonshot’s headline coding scores are vendor-run and should be treated as a ceiling; anchor on independent sources such as Artificial Analysis.

Who should use Kimi?

Great for:

• Researchers and analysts — long-document analysis and autonomous, cited research are Kimi’s core strengths
• Developers — open weights, a cheap API and a capable coding CLI (Kimi Code)
• Builders who want agents — OK Computer and Agent Swarm produce websites, slides and full-stack apps
• Cost-conscious users — a near-frontier model for $19/month, or free to self-host
• Anyone evaluating open weights — try the hosted app, then self-host the same model

Look elsewhere if:

• You handle sensitive or regulated data — use Le Chat (EU-hosted) or self-host an open-weight model so nothing leaves your infrastructure
• You want a free flagship — the DeepSeek app is free, with no paid tier
• You want an all-rounder — ChatGPT and Gemini cover images, video, voice and broad ecosystems better
• You need real-time voice — ChatGPT and Gemini are more natural

Historical timeline

Privacy and data handling

Kimi is a Chinese consumer service, and that is the central privacy consideration. Prompts and responses are processed and stored on servers in the People’s Republic of China, which brings the app within scope of China’s National Intelligence Law — the same concern that has led several governments to restrict DeepSeek and to scrutinise Qwen. Treat the app as unsuitable for sensitive, confidential or regulated data.

The practical escape hatch is the open weights. Because Kimi K2.6 is released under a modified MIT licence (which requires attribution for products exceeding 100 million monthly users or $20 million in monthly revenue), and Kimi K3 and K3.1 Pro under their own licences, you can run the same models yourself, or through a Western inference provider, so that no data reaches Moonshot’s servers. For most organisations, that — not the consumer app — is the right way to use Kimi for anything confidential.

Moonshot has not published the kind of consumer privacy certifications (SOC 2, ISO 27001, GDPR tooling) that Western providers detail, so retention periods and training-data practices for the app are not clearly specified. Assume consumer conversations may be used to improve the service.

Kimi vs alternatives

Quick comparison (July 2026, at Kimi K3’s launch)

Where Kimi wins

• Long-document analysis — purpose-built for reading across large document sets
• Agentic depth — OK Computer and Agent Swarm go further than most rivals’ agent features
• Capability — Kimi K3 ranked 4th of all models on Artificial Analysis’s index at its July launch, ahead of Claude Opus 4.8, and its weights are open
• Price — cheaper than Western subscriptions, and free to self-host

Where competitors win

• DeepSeek — entirely free, with a 1M-token context and no paid tier
• Qwen — free multimodal app with built-in image and video generation
• ChatGPT and Gemini — broader multimedia, real-time voice, larger ecosystems and US data hosting

For sensitive work, Le Chat is EU-hosted and GDPR-native, and any of the open-weight models can be self-hosted so nothing leaves your infrastructure.

Getting started

Web

1. Go to kimi.com
2. Sign in
3. Start chatting, or switch on OK Computer for agentic tasks

Mobile

• iOS: download from the App Store (requires iOS 14.0 or later)
• Android: download the Kimi app (requires Android 8.0 or higher)

Desktop

Kimi runs as a progressive web app on Windows, macOS and Linux from kimi.com; Moonshot also offers Kimi Work, a dedicated desktop agent app.

Tips for new users

Upload your documents first. Kimi’s long context shines when you give it the source material — drop in several PDFs or reports and ask for a cross-document summary or comparison.

Use OK Computer for deliverables. When you want a website, slide deck or data analysis rather than an answer, switch to agent mode and describe the outcome.

Self-host for anything sensitive. If data residency matters, run the open-weight Kimi models yourself or through a Western inference provider rather than using the app.

Frequently asked questions

Is Kimi free?

Yes. Kimi’s web and mobile apps are free to use with general rate limits, including chat, web search, file uploads and long-document analysis. The free tier runs a lighter model; the flagship model and the largest agent and research quotas require a paid plan starting at $19/month.

What is the latest Kimi model?

Kimi K3.1 Pro, released on 23 October 2026 in the app and on the API, with open weights since 4 November; Artificial Analysis scored it 55 on its Intelligence Index (v4.3.2) at launch. Its predecessor, Kimi K3, released on 16 July 2026, is a 2.8-trillion-parameter Mixture-of-Experts reasoning model with a 1-million-token context window and native vision. At launch, Artificial Analysis ranked it 4th of all models — ahead of Claude Opus 4.8, behind only Claude Fable 5 and GPT-5.6 Sol — then the strongest result any Chinese model had posted.

Is Kimi better than ChatGPT or Claude?

Not on top. At Kimi K3.1 Pro’s 23 October launch, Artificial Analysis scored it 55 on its Intelligence Index v4.3.2; Claude Opus 5.5 had scored 58 on v4.3 at its September launch. At Kimi K3’s July launch, on the Artificial Analysis Intelligence Index v4.1, it scored 57.1, above Claude Opus 4.8 (55.7) and below Claude Fable 5 (59.9) and GPT-5.6 Sol (58.9). It is genuinely competitive for coding, long-document work and agentic tasks, but ChatGPT and Claude remain stronger all-rounders with broader features, lower measured hallucination rates and US data hosting.

What is OK Computer in Kimi?

OK Computer is Kimi’s agent mode. It gives the model its own virtual computer to research, plan, design and build — producing multi-page websites, editable slides and data analyses, and processing up to one million rows of data in a single run. It is what turns Kimi from a chatbot into a tool that returns finished work.

Is Kimi safe to use with private data?

Use caution. The Kimi app stores data on servers in China and falls under China’s National Intelligence Law, so it is not suitable for sensitive, confidential or regulated information. Because the Kimi models are open-weight, the safer route for private work is to self-host one or run it through a Western inference provider so no data reaches Moonshot.

How much does Kimi cost?

The app is free with limits. Paid plans are Moderato ($19/month), Allegretto ($39/month), Allegro ($99/month) and Vivace ($199/month), with larger usage and agent quotas at each step; figures vary slightly by region and purchase channel. On the pay-as-you-go API, Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output, Kimi K3 $3 and $15 ($0.30 cached input), and the older K2.6 roughly $0.95 and $4.

Can I download and run Kimi myself?

Partly. The Kimi K2 model family, including K2.6, is released under a modified MIT licence, so those weights can be downloaded, fine-tuned and self-hosted (the licence requires attribution for products with over 100 million monthly users or $20 million in monthly revenue). Kimi K3.1 Pro’s weights followed on 4 November 2026 under the MIT-based Kimi K3.1 Pro License. Kimi K3’s weights shipped on 26 July 2026 under the Kimi K3 License (a permissive, modified-MIT-style licence) — a ~1.56TB MXFP4 checkpoint, the largest open model released at the time, though its size makes it far harder to self-host than the K2 line.

Who owns Kimi?

Kimi is made by Moonshot AI, a Beijing company founded in March 2023 by Yang Zhilin, Zhou Xinyu and Wu Yuxin. It is backed by Alibaba and Tencent, and in 2026 was reported to be considering a Hong Kong listing.

Verdict

Kimi is no longer just the specialist’s choice: with Kimi K3 ranked 4th of all models on Artificial Analysis’s index at its July launch, above Claude Opus 4.8, it became a frontier-adjacent assistant priced below its Western rivals. Long context and agentic depth remain the signature strengths, OK Computer and Agent Swarm produce real deliverables, and the open-weight Kimi models mean you are never fully locked in. For long-document analysis, autonomous research and agentic building, it is one of the best tools available.

Free tier: Generous for chat, web search and long-document work, but it runs a lighter model and caps the heaviest agent features.

Paid plans ($19–199/month): Moderato is the sweet spot — the flagship model (Kimi K3 at 256K context when Moonshot listed the plans in July 2026), Deep Research, the agent tools and Kimi Code for less than the cost of ChatGPT Plus; Allegretto adds the full 1M-token window.

The decisive caveat is jurisdiction. The app stores data in China, so for anything sensitive the right way to use Kimi is to self-host the open weights rather than the consumer app. Used that way, it is one of the most capable open AI systems you can run.

Official links