All Things AI · Season 3 · The New Models · Episode 3
Kimi K3 is a free chat assistant from Moonshot AI that reads a huge amount of context at once and can split a big research task into hundreds of smaller sub-tasks running at the same time, so a job that would take you an afternoon comes back in one pass.
What it is
Kimi K3 is a large model built for long, involved work: long documents, long research tasks, and reasoning that takes several steps to get through. It understands text, images, and video.
It reads over a million tokens of context in one go, which is enough room to hold a very long document, or a very long back-and-forth conversation, without losing track of the earlier parts.
It’s built by Moonshot AI, and Moonshot describes it as designed for long-horizon coding, knowledge work, and reasoning tasks, with agentic knowledge work that includes deep research delivered with interactive visualizations rather than a plain wall of text.
The Kimi app: a free chat assistant
The everyday way to use it is the Kimi app, at kimi.com or kimi.ai. You sign up with a Google account or a phone number, no credit card required, and the free tier gives you the K3 model along with a deep-research feature.
It also has an agent mode called ‘OK Computer’ that can build a multi-page website or an editable slide deck from a description, right inside the chat.
How you reach it
kimi.com or kimi.ai in a browser, or the Kimi app on Android. Sign-up is free and takes a minute.
Agent Swarm: splitting one job into hundreds of smaller ones
This is the feature that sets Kimi apart from a normal chat assistant. Agent Swarm can automatically coordinate a large group of AI sub-agents working on the same task at once, deploying up to 300 of them at the same time, and Moonshot’s own numbers put the whole thing at roughly four and a half times faster than one agent working through the same job step by step.
Moonshot’s own examples of what this is for: finding the top creators across a hundred different YouTube niches in one pass, building a hundred-page literature review out of forty separate documents, or getting a review of something from multiple different expert perspectives at once.
- One big research question, broken into many smaller ones, all answered at the same time.
- Reviewing a large stack of documents and pulling one combined summary out of all of them.
- Getting several different angles on the same question back in a single pass instead of asking one at a time.
Running Kimi K3 yourself, and the fine print on using it in your own business
Kimi K3’s weights are also released openly, so a business that wants to run it inside its own systems is free to do that under Moonshot’s own K3 license, and that license grants the right to use, copy, modify, merge, publish, distribute, and sell copies.
The license adds a couple of extra steps for very large companies specifically: a business bringing in more than $20 million in revenue over 12 months has to sign a separate agreement before selling K3 as a service, and a product with more than $20 million in monthly revenue or 100 million monthly users has to show the name ‘Kimi K3’ on screen. Neither of those thresholds applies to a solo agent or a small team using it for their own work, which is free and unrestricted.
The Kimi API
For anyone building Kimi into their own tools rather than using the chat app directly, there’s a straightforward API that works the same way as other popular AI APIs, which means most existing tools built for one of those can be pointed at Kimi by changing a single address.
It’s priced per million tokens, with a lower rate for tokens it’s already seen before through caching, which rewards a workflow that reuses the same reference material across many requests instead of starting fresh each time.
What to use it for
The profile is anything long, anything that needs several angles at once, or anything you’d normally break into a dozen separate searches yourself.
- A research question that has many parts, handed to Agent Swarm instead of worked through one search at a time.
- Reading and summarizing a large stack of your own documents in one sitting.
- A first draft of a website page or a slide deck from a plain description, through the OK Computer agent mode.
- Anything where the free tier’s context window, over a million tokens, means the whole document fits without cutting it into pieces.
How Kimi K3 was built to reach that speed
Under the hood, K3 uses a mixture-of-experts design, meaning it has a very large total number of parameters but only activates a smaller portion of them for any single answer, which is part of why it can move quickly despite its overall size. It also uses a hybrid attention approach Moonshot calls Kimi Delta Attention, built specifically to handle long context efficiently.
Since its release, K3 has debuted near the top of independent model leaderboards, including a top-three placement on evaluations that measure real business and research tasks, and a strong score on a benchmark that specifically measures how well a model can browse and research the web on its own.
Where it is weaker
Two honest limits.
The free tier is genuinely useful, but the extra-long chat mode, going all the way up to the full million-token window, is reserved for the paid tiers. For most day-to-day questions the free tier’s context is already generous.
And Agent Swarm’s real strength is breadth, coordinating many smaller angles on one job. For a single narrow, precise question, a normal chat response is usually faster than spinning up a swarm.
What to try in the first week
Two exercises, both built to show off the two different strengths.
First, take one long document you already have, something you’d normally read start to finish, and ask Kimi to summarize the whole thing at once. Notice that you never had to split it into chunks.
Then try the Agent Swarm feature on a research question with several distinct parts, something you’d normally run as five or six separate searches. Compare what came back against how long doing that by hand actually takes you.
How it sits next to the others
Against the other free chat assistants in this season, Kimi’s edge is the combination of a very large context window with Agent Swarm’s ability to fan a big job out into many smaller ones running together.
For a quick, single question, any of the free assistants in this season will do the job. For a genuinely large research task, Kimi’s swarm feature is the one built specifically for that shape of problem.
Questions agents actually type
Is Kimi K3 free?
Yes. The Kimi app at kimi.com or kimi.ai has a free tier that includes the K3 model, and sign-up needs only a Google account or a phone number, with no credit card required.
What is Kimi’s Agent Swarm?
A feature that automatically coordinates up to 300 AI sub-agents working on the same task at once, roughly four and a half times faster than working through the same job with one agent in sequence. It’s built for large research jobs with many smaller parts.
How much text can Kimi K3 read at once?
Over a million tokens of context, enough to hold a very long document or a very long conversation without losing track of the earlier parts.
Can Kimi build a website or a slide deck?
Yes, through its agent mode called OK Computer, which can create a multi-page website or an editable slide deck from a plain description inside the chat.
Who makes Kimi K3?
Moonshot AI. It’s reachable through the free Kimi app, and the model itself is also released with open weights that can be downloaded and run elsewhere, including inside a business’s own systems under Moonshot’s own license.
Does Kimi K3 cost money if I run it inside my own business?
Not for a solo agent or a small team. The license only adds extra steps for very large companies, above $20 million in yearly revenue or 100 million monthly users. Below that, using the open weights in your own work is free and unrestricted.
Book a 30-minute strategy call →
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.