Open models: Kimi, DeepSeek and Qwen

When we call a model open, what exactly is open

Three completely different things get called open, and until they are separated the rest of this page means nothing. First, open weights: the numbers the model learned are published as dozens of downloadable files and you can load those files on hardware you control. Second, open in the software sense: the training code and the training data are published too, and the license is one of the recognised ones such as Apache 2.0 or MIT. Third, something that is not open at all but is constantly confused with the other two: a model sold only through an API at a low price.

Nearly every model named on this page belongs to the first group only. The weights are there, the training data is not, and in several cases the license is not a standard license either but a text the vendor wrote and named after its own model. The gap between I can download it and I can do anything with it is exactly where a commercial project gets stuck, usually once the work is already finished.

Our position: ask which license this repository carries today, not whether this company is open source. The second question has no answer. One vendor can ship its small model under Apache 2.0 and its flagship under a conditional license, and that is precisely what it does.

Which models can you actually download

We read the table below straight from each vendor own repository on Hugging Face on 6 September 2026. The license column is what the model card records, not what the news coverage says.

Two things in this table are worth staring at. The first is that rows two and three come from the same company and do not share a license; the Qwen rows say the same thing, where the 27 billion model is Apache 2.0 and the flagship is not. The second is that the largest models on the list are exactly the ones with conditional licenses. That is not a coincidence: the conditions get written for the model with the most commercial value.

One note about names that Persian and Turkish listicles keep getting wrong: the name on a leaderboard is often not the name of a model. Kimi K3 Max and GLM-5.2 Max are not separate models in the vendor documentation, where max is a reasoning effort level. So go looking for the repository, not for the string printed in a ranking table.

How to read the license: three clauses that rhyme

We put the three bespoke licenses on this list side by side and read all three. They are structured almost identically: an extremely permissive MIT-style grant first, then two conditions that only switch on above a revenue threshold, and finally an explicit carve-out saying internal use is exempt from both.

In the Kimi K3 license, the first condition reads: if your business is model as a service, meaning you let a third party exercise meaningful control over inputs and parameters, and your aggregate revenue over any consecutive twelve months exceeds twenty million dollars, you must enter a separate agreement with Moonshot before any commercial use. The second says a product with more than one hundred million monthly active users or more than twenty million dollars in monthly revenue must display the name Kimi K3 on its interface. Clause four then states plainly that neither applies to internal use.

The GLM-5.3 license follows the same template with a far higher threshold: ten billion dollars of revenue over twelve months, and instead of an agreement it requires passing a Z.ai security review. The Qwen3.8-Max license sets fifty million dollars and adds a second category called an AI work assistant, meaning a standalone product built primarily for coding or office productivity.

Now the practical conclusion, and this is where most coverage either frightens people or flattens the whole thing. For a company with a few dozen employees, all three thresholds above are irrelevant. If you run the model on your own server to do your own work or your client work, you are effectively holding an MIT-shaped license. The license that actually bites is somewhere else: CC BY-NC. That one forbids commercial use at any size, and it sits on models nobody expects, such as Cohere multilingual translation model Aya Expanse 32B. Put that inside a client project and the size of your company does not matter.

Read the model card before you build on it

The model card is the front page of the repository, and five things on it can be read in fifteen minutes and will save weeks later. First the LICENSE file inside the repository itself, not the badge at the top of the page. For three of the models on this list that badge only says other, and the real text is inside the file. Second, whether there are weight files at all: a repository with no safetensors files is an explainer page, not a runnable model.

Third the parameter count, which decides what this model will even start on, and which the next section is about. Fourth whether the version is base or instruct: a base model does not follow instructions, and if you expect chatbot behaviour from one you will conclude the model is broken. Fifth the date the repository last changed, because as the table above showed, the license changes between two versions.

And one trap that belongs specifically to real work: quantised re-uploads. The community publishes four bit and eight bit versions of these models in its own repositories, and those are the ones that fit on ordinary hardware. Those repositories do not always carry the license of the original, and some record no license at all. Build on one of them and what you hand a client has an unknown license. Hugging Face own guidance on repository licenses says the license is declared per repository, and that single sentence is the whole story.

Open weights mean a supported-countries list stops applying

This is an architecture fact, not a workaround. An API has an account, an account has a country, the country is either on the supported list or it is not, and payment needs a card. A weights file has none of that: it is a set of files loaded on a machine you chose yourself. So does this model work from Iran is not even the right question about open weights. The right question is where that machine is and how much memory it has.

We do not write circumvention advice on this site and we do not sell accounts, which is what our AI from Iran page already says. All we are saying here is that open weights are a different access layer, not a trick. And we state its two honest limits too: to download it you still have to reach the repository, and to run it you need hardware.

Hardware is where most optimistic threads end. A rule of thumb you can work out yourself: at 16 bit precision every parameter takes two bytes. For the 27.8 billion parameter Qwen model that is roughly 55 gigabytes for the weights alone, before anything else. For GLM-5.3 at 753 billion parameters it is roughly one and a half terabytes. Community quantised builds cut that figure several times over and they are what makes local use possible at all, but we have not run any of them on our own hardware and we will not invent a number.

So the honest conclusion is this: the small models in these families run on an ordinary rented server, and the flagship models on that same list do not. Anyone who says DeepSeek is open source so you can run it for free has probably never loaded 1.65 trillion parameters anywhere.

Which one to pick for real work

There are three cases and the answer differs in each. If you only want a first-rate answer more cheaply and local execution is not your issue, open weights are of no use to you at all: buy the same models from their own API and look at a comparison number. If you hold data that must not leave your server, then open weights are the only route, and you should start from a small model rather than a flagship. And if you want to put a model inside a product you sell to a client, read the license first and write code second.

We are not telling you which model is better here, because that answer changes week by week and this is not its place. Its place is the AI reference section of this site, where every row carries a read date and a source. What this lesson gives you is knowing what the license column means when you read that table.

Let us also draw a boundary: all of this is about language models. Image and audio models carry different and sometimes stricter licenses, and models whose output goes straight to a client raise a further question covered elsewhere in this track: who legally owns what the model produced.