Kimi is Moonshot AI's chat assistant, built by the Beijing company whose open-weight K2 and K3 models spent much of 2026 near the top of the coding and agent benchmarks. You use it the way you'd use ChatGPT or Claude: ask a question, paste in a 200-page contract, hand it a repository and ask for a refactor. Two things set it apart — a 1-million-token context window on the K3 model, and Agent Swarm, a mode that spins up a crowd of sub-agents to work through large jobs in parallel. It's aimed at people who do heavy document and code work and want a capable free tier while they weigh up paying.
Key Features
- 1M-token context on the K3 model, so you can load a whole codebase or a stack of reports at once and it keeps cross-references straight instead of losing track of file 3 by the time it reads file 12.
- Agent Swarm: one orchestrator hands work to as many as 300 sub-agents running at the same time, up to 4,000 tool calls per task. Moonshot claims roughly 4.5x faster completion on jobs that split into parallel chunks, like batch research or reviewing 100-plus documents.
- A free tier you can genuinely live on — unlimited basic chat, file uploads, live web search and image analysis, with conversations on the K2.6 model not spending your monthly agent credits.
- Kimi Work, a free desktop agent for macOS and Windows that mounts your local folders, drives a browser on its own, and runs Python in the background to turn live data into finished documents.
- Office files in, office files out — Kimi edits and returns Word, Excel, PowerPoint and PDF rather than just describing the changes, and Kimi Code adds a command-line coding agent with VS Code and Zed support.
- Sync across web, iOS, Android and desktop, so a chat you start on your phone is waiting on your laptop.
Screenshots
Why You'll Love Kimi
The long-context handling is the real thing, not a spec-sheet number
Plenty of models advertise a big context window and then quietly truncate the middle. Independent testers have dumped 40,000-line codebases into K3 and asked it to refactor a function referenced across a dozen files, and it traced every call site without the "remind me what that file does" loops you get when a model loses the thread. For legal, research and finance work, where the right answer depends on page 210 lining up with page 4, that reliability matters more than a couple of benchmark points.
Agent Swarm is impressive when the task actually fits it
Parallelism only helps when the work splits cleanly: comparing 200 vendor pages, pulling figures from a pile of filings, drafting a long report section by section. Point it at that kind of job and an afternoon turns into a coffee break. Point it at a quick question and it's overkill that burns through credits fast. It's also paid-tier only, so treat it as a heavy tool you reach for deliberately, not the default.
Plans and pricing
The free Adagio tier covers everyday chat, file uploads and web search, and it doesn't nag — K2.6 conversations don't even touch your credit balance. K3 and Agent Swarm need a paid plan: $19/month for Moderato (60 agent credits, 25 swarm runs), $39 for Allegretto, $99 for Allegro and $199 for Vivace, each around 20% cheaper on annual billing. Credits reset monthly and don't roll over. One thing to know before you get attached: after K3 launched in July 2026, demand overran Moonshot's GPUs and every paid tier sold out, with new subscriptions reopening only in batches. If you're outside China, check that you can actually buy a plan before you build a workflow around Swarm. API access is separate, billed per token, with K3 at roughly $3 per million input tokens and $15 per million output.
Category: AI Assistants
Platform: Web, iOS, Android, macOS, Windows
Price: Free tier (Adagio); paid plans $19–$199/month
Frequently Asked Questions
Yes. The Adagio tier is free with no time limit and gives you unlimited basic chat on the K2.6 model, file uploads, live web search and image analysis, plus the free Kimi Work desktop agent. What's gated is the K3 flagship model and Agent Swarm, which need a paid plan starting at $19/month, along with a monthly pool of agent credits for heavier tasks. Note that new paid subscriptions have been capacity-limited since mid-2026.
No. Every response is generated on Moonshot's servers, so you need a connection whether you're on the web, the mobile apps or the desktop app — there's no on-device mode. The K2 and K3 model weights are open, so a technical user can self-host them separately, but the Kimi app itself is online-only. One privacy point worth knowing: your data is processed on infrastructure in China.
On Android the core app only needs network access. Everything sensitive is optional and requested in context: microphone for voice input, camera and gallery access for taking photos, uploading images or scanning the QR login code, storage for file uploads, and notifications so you know when an agent task finishes. Decline any of them and the rest of the app keeps working normally.