What is Kimi, and what open weights really means
Kimi is Moonshot AI's model family. Read what its licence actually says, then work out whether open weights change anything for you.
You have most likely met the name Kimi in a headline about Kimi K3, published in July 2026, and that headline almost certainly used the phrase "open weights" without saying what it covers. Kimi is a family of AI models built by Moonshot AI, a Chinese lab that also runs its own Kimi chat app, browser extension and desktop clients.
The phrase does a lot of quiet work. It attaches to Qwen, GLM, DeepSeek and a dozen other names, and it gets used as a synonym for open source. Those are different things. Kimi K3 is the clearest illustration going: Moonshot published a licence you can read in four minutes and a download almost nobody reading this could run. Work through these five checks and you will know what the label means for you, and for every other model carrying it.
1. Open the licence file before you trust the announcement
Kimi K3's weights sit on Hugging Face under a document headed "Kimi K3 License", a bespoke text rather than one of the standard licences.
Most of it is generous. It grants any person, free of charge, the right to use, copy, modify, merge, publish, distribute, sublicense and sell the software, and to run, deploy and fine-tune it. Then come two conditions that no standard licence contains.
The first is a revenue gate on serving the model to other people. If a licensee "operates a Model as a Service business, and the aggregate revenue of the Licensee and its affiliates exceeds 20 million US dollars (or the equivalent in other currencies) in total over any consecutive 12 months, the Licensee must enter into a separate agreement with Moonshot AI before using the Software or its derivative works for any commercial purpose."
The second is a billboard clause. Past 100 million monthly active users or 20 million US dollars in monthly revenue, "Kimi K3" must be prominently displayed on the user interface of such product or service.
The two older Kimi models we carry are governed by a different file headed "Modified MIT License", which keeps the naming requirement and drops the revenue gate entirely.
Check the model, not the family. One vendor can ship three licences in one product line, and Moonshot has. "Kimi is open" describes one specific checkpoint on one specific day, and nothing wider.
2. What was left out of the box
A download page showing 1.56 terabytes of files looks like everything. It is not.
The K3 repository holds 119 files. Ninety-six of them are weight shards, the frozen numbers the model learned during training. The rest are configuration, the tokenizer, and the Python that loads and runs it. There is no training data. There is no record of the training run you could repeat.
That gap is the whole distinction. Open weights means the finished parameters are downloadable under some licence. Open source, in the sense the Open Source Initiative defines it, means you get the preferred form for making modifications and the freedom to pass those rights on untouched. With weights alone you can run the model and adjust it at the edges. You cannot rebuild it, audit what went into it, or verify a claim about what it was trained on.
Why this is not pedantry. The Open Source Definition requires that "the rights attached to the program must apply to all to whom the program is redistributed without the need for execution of an additional license by those parties." Kimi K3's revenue gate requires exactly such an additional agreement from anyone large enough. Artificial Analysis now treats openness as a spectrum for that reason, on an Openness Index that "assesses model openness on a 0 to 100 normalized scale (higher is more open)".
3. Work out who could physically run it
Here is where "open" stops meaning what it sounds like.
That 1.56 terabytes is the download, before you have run a single word through it. The inference recipe Moonshot's own model card points to states the requirement plainly: "At least 8x GB300. Multi-node for real production traffic." Those are data-centre accelerators, bought eight at a time.
So the practical reading of "open" here is that anyone with a server hall can run it and you cannot. The download link is real and genuinely useful, just not to you personally, and no amount of laptop changes that.
Do not read "open" as "private" either. A downloadable model does not mean the copy answering you is running on your own machine. It is running on somebody's servers, and which somebody is a separate question with a separate answer.
4. Sort what moves from what does not
Four things get credited to published weights. Three of them hold up.
Someone other than the vendor can run the model. This is the real consequence. Independent hosts can serve Kimi, which is why you reach it from apps Moonshot has nothing to do with. In our case the three Kimi entries can only be served by hosts on an approved list, and Moonshot's own endpoint is not on it.
The model can outlive the company's interest in it. A proprietary model can be switched off and become genuinely unavailable. A published checkpoint stays on disk somewhere, and anyone with the hardware can keep serving it.
A business or a self-hoster gets options, such as pinning one version while the vendor moves on. That matters enormously to a few readers and not at all to most.
What the answers look like on your screen. This one does not change. A licence file has no effect on how the model writes, how fast it replies, or how often it is wrong.
Weights on a download page say nothing about how your conversation is handled. That part reads the same for every model we carry: we do not train on your conversations, some providers may depending on which model you pick, and where we know of a difference, that model's info card in the app carries a notice.
5. Check the working limits before you pick a Kimi model
We carry three Kimi models among the 59 in the app, all of them on the paid plans: K3, K2.6 and K2 Thinking. What matters day to day is what they will and will not accept.
No web search on any of them. There is nothing to switch on. If your question needs today's news, pick something else.
No PDF attachments. All three decline PDF uploads, so the attach will be blocked before it starts.
Images in, on two of the three. K3 and K2.6 accept image uploads. K2 Thinking takes text only.
Reasoning is always on. Every Kimi entry shows its working, and none of them lets you switch that off.
K3 advertises a one-million-token context window. Read that as a ceiling rather than a promise; our piece on context windows explains why the two come apart.
Expect it to take its time. Artificial Analysis, which re-benchmarks continuously, currently rates K3 among the leading models for intelligence and describes it as "notably slow and somewhat verbose". That standing moves as the page re-runs, so read the direction rather than the position. The intelligence is why people write about it. The waiting is what you notice.
The short answer
For your actual chat sessions, Kimi's licence changes nothing. It does not make the answers better or safer, and any article implying otherwise has confused a legal document with a product feature.
What it changes is who is allowed to hold the model: why Kimi is reachable from somewhere other than Moonshot's own app, and why it stays reachable if Moonshot loses interest. That is a fact about the supply chain, and it lands nowhere near your Tuesday afternoon.
All three Kimi entries sit behind the same subscription as the rest of the catalogue, so you can try them without a separate Kimi subscription.