What is Kimi and how it can help you be more productive

Moonshot AI assistant guide · Updated September 2026

Kimi isn’t just “China’s long-context chatbot” anymore. What started in October 2023 as Moonshot AI’s consumer assistant with a then-record context window has grown into a full open-weight platform: frontier-scale models, agentic products, native multimodal understanding, and the largest open-weight release ever. On July 16, 2026 Moonshot shipped Kimi K3 – a 2.8-trillion-parameter sparse MoE flagship with a 1-million-token context window and native vision – followed by full open weights on July 27.

In this guide we explain what Kimi is in 2026, how the K3 / K2.7 / K2.6 / K2.5 family works, what Agent Swarm, Kimi Code, and OK Computer can do, what the plans cost, and how to use it well.

Oct 2023First public release, long-context Kimi chatbot
Kimi K3Current flagship · 2.8T open-weight MoE · Jul 16, 2026
Swarm · Code · ClawAgentic modes that finish tasks for you
01 · What it is

What is Kimi?

Kimi is an AI assistant and family of large language models developed by Moonshot AI, a Beijing lab founded in March 2023 by Yang Zhilin (CEO), Zhou Xinyu, and Wu Yuxin – classmates at Tsinghua University. Yang earned his PhD at Carnegie Mellon and previously worked at Google Brain and Meta, and built Moonshot around one idea: long context is a first-class feature, not an afterthought. The company name nods to Pink Floyd’s The Dark Side of the Moon.

The Kimi chatbot entered closed beta in October 2023 and opened to the public that November with 128,000 tokens of lossless context – at the time, the first model that could actually accept that much in one go. Unlike closed US labs, Kimi’s defining trait since then has been open weights plus aggressive long-context engineering plus agentic coding. Most Kimi models ship with downloadable weights on Hugging Face so you can self-host, fine-tune, and keep data on your own infrastructure.

Today Kimi is not just a chat window. Depending on the plan and the task it can:

  • Answer questions & explain concepts
  • Write, rewrite, summarize & translate
  • Search the live web with citations
  • Reason step by step with thinking modes
  • Analyze files, spreadsheets, PDFs, images & video
  • Write, debug, test & ship code
  • Coordinate parallel agents (Agent Swarm)
  • Generate slides, sites & documents
  • Hold real-time voice conversations
  • Run locally on your own hardware with open weights
Kimi – the models

Kimi’s current generation is K3, unveiled July 16, 2026 and released in full weights July 27: a 2.8-trillion-parameter MoE with 896 experts (16 active per token, ~104B active), a 1M-token window, and native vision. It sits above the K2 family – K2.5 (Jan 2026), K2.6 (Apr 2026), K2.7 Code (Jun 2026) – which remain widely used for everyday and coding workloads.

Kimi – the products

You meet Kimi through the consumer chat (kimi.com, mobile apps), Kimi Work (desktop agent), Kimi Code (CLI and coding agent), Agent Swarm (parallel sub-agents), OK Computer (sites and slides from prompts), the Kimi API, and downloadable open weights.

02 · The model family

Kimi K3 and the K2 family

K3 is a 2.8-trillion-parameter Mixture-of-Experts model with 896 experts, of which 16 are active per token. It is built on two architectural innovations Moonshot developed in-house: Kimi Delta Attention (KDA), which cuts memory use and speeds up long-context generation, and Attention Residuals, which stabilize training across depth. K3 debuted near the top of independent leaderboards – behind only the strongest closed models on composite indexes, while leading some coding and front-end arenas.

Frontier open-weight · Jul 16, 2026 (weights Jul 27)

Kimi K3

2.8T total / ~104B active MoE, 1M-token context, native vision (text + image + video). Designed for long-horizon coding, agentic knowledge work, and reasoning. Open weights under the Kimi K3 License – the largest open-weight release ever at 1.56 TB.

Maximum capability · open weightsFlagship
Specialized coding · Jun 2026

K2.7 Code

A code-specialized successor in the K2 line with a 256K context window and multi-step tool invocation. The specialist endpoint for pure coding workloads at a lower price than K3.

Best coding value · open weights
Workhorse · Apr 2026

Kimi K2.6

~1T total / 32B active MoE, 256K context, built for long-horizon coding and autonomous execution. Powers Agent Swarm orchestration and Kimi Code effectively – the practical default before K3.

Best value everyday model
Multimodal pioneer · Jan 2026

Kimi K2.5

1T-parameter MoE with native vision via the MoonViT encoder and 256K context. First Kimi with video understanding and Agent Swarm – up to 100 parallel sub-agents at launch.

First native multimodal · Swarm debut

Two notes on the wider lineup. First, K3’s weights are “open” with a caveat: the Kimi K3 License requires large inference providers to negotiate separately and large products to display attribution – the older K2 family ships under a more permissive modified MIT license. Second, behind K2.5 sit K2 (Jul 2025, the original trillion-parameter open release), K1.5 (Jan 2025, reasoning), and the Kimi-VL vision line.

03 · How it works

How does Kimi work?

Kimi models are Mixture-of-Experts transformers trained natively on text, images, audio, video, and code together. Only a fraction of total parameters activate per token – that is how K3 can be 2.8T total but cost ~104B active per forward pass. A thinking-level setting controls how much the model reasons step by step before answering: trivial prompts get a fast response, hard ones trigger deliberate multi-step reasoning. K3 runs at maximum thinking effort by default.

Modern Kimi goes beyond next-token prediction by combining the model with tools and, in Swarm mode, parallel sub-agents. A typical request flows through several stages:

  • 1

    Understand the request

    Reads the prompt, conversation history, uploaded files/images/video, and any connected tools, memory, or MCP servers.

  • 2

    Reason through the problem

    For complex tasks, the thinking level lets it work through the problem step by step instead of answering in one pass. K3 always thinks; effort accepts low, high, and max.

  • 3

    Decide which tools to use

    Web search, code execution, file analysis, image or video generation, or hand-off to sub-agents in Swarm mode.

  • 4

    Generate the response or take the action

    Produces the answer, the document, the code, or the finished deliverable – sometimes after sub-agents have worked in parallel for minutes or hours.

  • 5

    Refine with follow-ups

    Context compaction, checkpoints, and plan mode keep long sessions coherent across hundreds of thousands of lines of code or dozens of papers.

04 · What you can do

What can you do with Kimi?

Research and information

Web search grounds answers in live results with citations. Deep research goes further: K3 Swarm plans a multi-step approach, deploys sub-agents to browse sources in parallel, and returns a cited, multi-page report.

Writing and editing

Brainstorm, outline, draft, rewrite, translate, and summarize. Kimi handles long-form documents – up to entire books in a single 1M-token session – and keeps consistency across hundreds of pages.

Coding and DevOps

K3 sustains long engineering sessions across massive repositories – writing, debugging, refactoring, testing, and reviewing in most languages. Kimi Code brings it to the terminal; K2.7 Code is the cheaper specialist endpoint.

Long-document analysis

The original superpower. Upload 100+ page contracts, reports, or a full codebase. Kimi started at 128K lossless in 2023 and now holds 1M tokens in K3, with context caching making repeat queries cheap.

Data analysis

Upload spreadsheets, CSVs, PDFs, and datasets. Kimi cleans data, computes metrics, finds trends, and renders charts rather than only describing them in prose.

Images and video

Kimi is natively multimodal: upload screenshots, diagrams, photos, and documents to analyze, and generate or edit images and video. Vision-in-the-loop means K3 can write code directly from UI mockups and video walkthroughs.

Self-hosting and control

Every K2 onward ships open weights. K2 runs on modest hardware; K3’s 1.56 TB checkpoint needs a multi-GPU server. Self-hosting removes per-token costs – and the data-residency question – entirely.

Agentic workflows

OK Computer creates multi-page websites and editable slides from simple prompts. Kimi Code is a command-line agent for terminal workflows. Agent Swarm coordinates up to 100 specialized sub-agents in parallel.

05 · From chatbot to agent

The agentic shift: Swarm, Code, Claw, and OK Computer

The biggest change since 2023 is not that answers got better – it is that Kimi moved from conversation toward execution. Instead of “you ask, it answers,” the workflow is increasingly “you define a goal, Kimi researches, analyzes, uses tools, and comes back with something finished.”

01

Agent Swarm

Introduced with K2.5 and expanded under K2.6/K3: instead of one agent deciding everything, many specialized sub-agents coordinate – frontend, backend, testing, docs – handing off actions automatically. Up to 100 parallel agents for full-stack generation, cutting execution time by ~4.5× on parallelizable work.

Parallel multi-agent coordination

02

Kimi Code

A terminal agent for real engineering: reads a repository, edits files, writes and runs tests, and manages Git for long-horizon tasks. Works through the CLI and integrates with editors – the surface where K3’s sustained-session strength shows most.

CLI · Editor integrations · Long-horizon sessions

03

OK Computer & Kimi Claw

OK Computer (Sep 2025) turns prompts into multi-page websites and editable slides. Kimi Claw controls the browser and runs an always-online assistant in the cloud. Together they cover the “operate the computer for me” half of agentic work.

Sites & slides · Browser control · Cloud agents

Moonshot’s path here has been distinctive. While Western labs built agentic features into chat apps, Moonshot open-sourced the weights and let the ecosystem build around them – Cursor, Harvey, and others ship products on Kimi models. The strategy is fewer surfaces that each do more, backed by weights anyone can download and modify.

06 · Key moments

Kimi timeline: 2023–2026

Mar 2023

Moonshot AI is founded in Beijing by Yang Zhilin, Zhou Xinyu, and Wu Yuxin.

Oct–Nov 2023

Kimi chatbot enters closed beta in October and opens to the public in November with 128K lossless context – a record at the time.

Jul 2024

Context caching enters public beta, making repeated long-context workloads cheap.

Oct 2024

Kimi Explore Edition (now Kimi Researcher) launches globally with autonomous search; monthly active users later exceed 36 million.

Jan 2025

Kimi K1.5 reasoning model ships, claiming parity with contemporary o1-class models on math, coding, and multimodal tasks.

Jul 2025

Kimi K2 released – the first trillion-parameter open-weight MoE (1T total / 32B active) with same-day weights.

Sep 2025

OK Computer agentic feature added; K2-Instruct scores strongly on SWE-bench Verified.

Jan 2026

Kimi K2.5 ships – native multimodality plus Agent Swarm v1 for parallel multi-agent work.

Apr 2026

Kimi K2.6 launches as a general-purpose multimodal model with improved long-context code generation and reasoning.

Jun 2026

Kimi K2.7 Code ships as a specialized coding model with 256K context and multi-step tool invocation.

Jul 16, 2026

Kimi K3 unveiled: 2.8T parameters, 896 experts, KDA and Attention Residuals, 1M-token context, native multimodal – debuting near the top of independent leaderboards.

Jul 27, 2026

Full K3 model weights released under the Kimi K3 License – the largest open-weight model ever published at 1.56 TB.

07 · Staying productive

How to use Kimi well

Be specific

State audience, goal, format, and constraints up front. “Turn this into five board-ready bullets with one evidence line each” beats “summarize this.”

Match model to job

Everyday or high-volume → K2.5 or K2.6. Pure coding → K2.7 Code. Serious long-horizon coding, research, or agent work → K3. Use Swarm when parallel exploration helps.

Leverage the cache

K3’s prompt caching drops input cost by ~90% when you reuse the same opening context. Keep your system prompt stable for coding agents and RAG pipelines.

The prompt pattern that works

Role / Context:You are a senior frontend developer.Task:Reconstruct this website from the 90-second video walkthrough attached.Format:React components, responsive design, accessibility tags included.

Give clear role + task + format, attach real files or codebase context, and iterate. For long-horizon work, use Kimi Code in plan mode and let Agent Swarm decompose: one agent for archaeology across a large repo, one for implementation, one for tests.

08 · The window

How large is the context window?

Kimi K3 supports a standard 1-million-token context window – roughly 750,000 words – with flat per-token pricing across the entire window. That is enough to hold an entire codebase, a book-length document, or hours of video transcripts in a single prompt. The K2 family supports 256K tokens (about 192,000 words), and the legacy K2 supports 128K.

Kimi K3 – 1,048,576 tokens≈ 750,000 words
Kimi K2.6 / K2.7 Code / K2.5 – 262,144 tokens≈ 192,000 words
Kimi K2 (legacy) – 128,000 tokens≈ 96,000 words
Kimi (2023 launch) – 128,000 tokens≈ 96,000 words

Practical rule of thumb: a token is roughly 0.75 English words, and about 1.5 tokens per word once you factor in punctuation and spacing. Deep reasoning and Agent Swarm consume part of that budget, and attention is not perfectly uniform across the full window.

09 · Plans and pricing

What does Kimi cost?

PlanPriceModel accessLimits
Adagio$0K3 chat, file upload, web search~6 agent credits, no Swarm / Code / Claw
Moderato$19/moK3 + K2.660 agent credits, 2 concurrent tasks, 25 Swarm runs
Allegretto$39/moK3 + K2.6150 agent credits, 50 Swarm runs, Kimi Claw
Allegro$99/moK3 + K2.6, heavy agent work360 agent credits, 4 concurrent tasks, 120 Swarm runs
Vivace$199/moMaximum everything720 agent credits, 240 Swarm runs, 8 sub-agents
API · Kimi K3$3.00 / $15.00per 1M tokens in/out, 1M context$0.30 cached input · frontier pricing
API · K2.7 Code$0.95 / $4.00per 1M tokens in/out, 256K contextSpecialist coding endpoint
API · K2.6$0.95 / $4.00per 1M tokens in/out, 256K contextGeneral-purpose workhorse
API · K2.5$0.60 / $3.00per 1M tokens in/out, 256K contextMultimodal value tier

Annual billing saves roughly 20%, and the app and API are separate wallets. K3 undercuts comparable closed flagships on list price, and measured cost per completed task is often lower. The chart above reflects September 2026 – check Moonshot’s official pages for current rates.

10 · FAQ

Frequently asked questions

Is Kimi free?

Yes – the Adagio tier gives K3 chat with usage caps, and the open weights let you self-host at no per-token cost. Paid plans start around $19/month for meaningful agent work.

Is Kimi really open-source?

K3’s weights are downloadable, but the Kimi K3 License is not a pure open-source license: large inference providers must negotiate separately and large products must display attribution. The K2 family ships under a more permissive modified MIT license.

Can I run Kimi K3 locally?

Technically yes, practically only at scale: the 1.56 TB checkpoint needs roughly 16 high-end GPUs to serve. For most teams the API is cheaper; K2.6 runs on far more modest hardware.

How does Kimi compare to ChatGPT, Claude, or Gemini?

K3 competes at the frontier – ranking just behind the strongest closed models on composite indexes while leading some coding and front-end arenas. Its edge is open weights, the 1M window, Swarm parallelism, and lower cost per task; closed rivals lead on ecosystem breadth and some reasoning benchmarks.

Does Kimi have an app?

Yes – web at kimi.com, iOS and Android apps, Kimi Work for desktop, Kimi Code for the terminal, and Kimi Claw for always-online cloud agents.

Heads-up:Kimi is a fast-moving product. Plan names, model names, and prices change frequently – the details above reflect September 2026 (K3 released July 16, weights July 27). Check Moonshot’s official Kimi site and API docs for current availability and limits.

Related reading