Kimi K3, Mixture of Experts and the rise of open AI models

Last week the Chinese company Moonshot AI released a new model, Kimi K3. With 2.8 trillion parameters it is the largest open model ever built, and in benchmarks it scores among the very best in the world, just behind the top models from OpenAI and Anthropic. On tasks such as coding it even wins in some tests.

The real surprise is not only that Kimi K3 was made by a startup that is just three years old, but also in how the AI model works: it never switches on all of its parameters at once. For each piece of text it uses only about 1.8% of its size. In AI jargon we call this a Mixture of Experts model, and it is what keeps such a large model affordable for the end user. At a task level it is almost three times cheaper than Claude Fable 5. A good moment to explain how a Mixture of Experts model pulls that off.

What is Kimi K3 and why is it big news

Kimi K3 came out on 16 July 2026 and immediately became one of the talking points of the summer. The model can take in a million tokens at once, which we call the context window. A token is a small piece of text, roughly part of a word, so a million tokens comes down to a large codebase or around 1,500 pages of A4 in one go. And it can keep working on a single job on its own for a long time. That makes it less a chatbot and more a digital colleague.

How good is it really? Impressive, but not the outright number one. On an important coding benchmark it comes out on top, above Claude and GPT. In the broad intelligence tests it lands around third or fourth place, depending on the test. One caveat: many of the best figures come from Moonshot's own test setup, and those are not always a fair comparison. But there is no doubt that it belongs among the best in the world.

How can a three-year-old startup do this?

Moonshot AI was founded in 2023 by Yang Zhilin, a leading researcher who previously worked at Google Brain and Meta AI. Their chatbot Kimi is one of the most popular AI apps in China.

They are growing fast. At the end of 2025 the company was worth around 4 billion dollars, and in May 2026 they closed a round of 2 billion dollars at a valuation of more than 20 billion, with Tencent and Alibaba among the backers. By now there is even talk of 30 billion. In terms of valuation that is not comparable to OpenAI or Anthropic, and sits closer to the French company Mistral. But this could well be a platform that takes on Silicon Valley.

The chip restrictions that the United States imposes on China may slow Chinese development, but at the same time they seem to do something else: they force Chinese companies to build smarter, more efficient models. Kimi K3 is a textbook example of this, and that brings us to the heart of the matter.

What is a Kimi K3 Mixture of Experts model?

Kimi K3 has 2.8 trillion parameters. Parameters are the control dials of an AI model, the place where everything it has learned is stored. More parameters roughly means more knowledge and more skill. But there is a catch: the more parameters a model uses at once, the more computing power, electricity and money each question costs. A giant of 2.8 trillion parameters that switched everything on for every word would be unaffordable.

This is where the Mixture of Experts, or MoE for short, comes in. Think of it as a large consultancy with 896 specialised teams. For each piece of text (token) that comes in, the system selects only the 16 teams that know most about it at that moment. The other 880 stay quietly in their seats. So you have the knowledge of the whole firm on hand, but for each question you only pay the teams you actually put to work.

Kimi K3 did not invent this approach. MoE is now the standard behind just about every leading model. The idea goes back to research from the 1990s and has been widely adopted in recent years, among others by DeepSeek, Mistral and Soofi S. The makers of the closed models from OpenAI and Anthropic do not reveal their design, but it is generally assumed that they work this way too, simply because it is the only way to keep a model this large and still affordable. What makes Kimi K3 special is not the trick itself, but the scale at which it applies it: 896 experts is a lot, and exceptionally large for an open AI model.

So how much of the model does it actually use? Of those 2.8 trillion parameters, only around 50 billion are active per token. That is less than 2%. This lets the makers grow a model almost endlessly in knowledge, without the bill for each question growing just as fast. It is the key technique that allows the newest generation of models to be both large and affordable at the same time.

Open-weight is not the same as open source

Kimi K3 is an open-weight model, not open source. From 27 July 2026 you should be able to download the trained weights to run the model yourself. Those weights are the result of the training, the dials as they have been set. But the full recipe underneath, so exactly how it was trained and on which data, is not included. That, by the way, is true for all powerful AI models.

And running it yourself is easier said than done. For the full version Moonshot recommends 64 AI accelerators, in other words specialised computing chips. That is far more than an average company will buy for itself. For now you can only reach the model through online services and the API (the programming interface). That runs via an entity in Singapore, so for sensitive company or personal data it is not something you can simply use.

The opportunity for Europe

It is precisely that open-weight character that offers Europe an opportunity. Once the weights are released, a European hosting party can download the model and offer it as a service under the right licence. It then remains a Chinese model, but it runs here and the data stays here, under European rules.

That touches on a bigger game, what we call "soft power". Whoever supplies the models, standards and infrastructure gains influence over the global AI ecosystem. I have written before that AI has long stopped being only about efficiency, and is now about geopolitics. Last Friday, at the major AI summit in Shanghai, Xi Jinping positioned China as the party that makes AI accessible to the whole world, in contrast to the dominance of the largely closed American frontier models.

But Europe is building too. In exactly the same week, on 13 July, a German consortium came out with Soofi S, an open model trained in Munich on the cloud of Deutsche Telekom (https://arxiv.org/abs/2607.09424). It is much smaller than Kimi K3, with about 90 times fewer parameters, and it is mainly good at German and English. But it is built entirely for European sovereignty. It is still a base model though, not a ready-made digital colleague.

Three tips for companies looking to get started with AI

1. First, take these kinds of models seriously in your comparisons, especially for coding, web development and document research. And look not only at the price per token, but at the cost of a completed task. A model that is slightly more expensive per token but finishes the job right in one go can work out cheaper in the end.

2. Second, do not build your business process around a single supplier. There is freedom of choice: one model for writing, another for code, and a European or local model for sensitive data. Stay flexible.

3. Third, do not assume that "open" means you will just run it yourself. Over the coming years I expect many European providers to host open models as a ready-made AI service in European data centres. That capacity will be scarce for a while, because of the limited computing power in European data centres. But all in all this is mainly good news for European companies, because more strong open models means more choice and less dependence.

Frequently asked questions

What is a Mixture of Experts model in short?

It is an AI model that is divided into many specialised parts, "experts", of which only a few are switched on at once for each piece of text (a token). This way the model holds a huge amount of knowledge while the computing costs stay limited. Interested in the technical stuff? Read here.

How much of Kimi K3 is actually switched on per question?

Of the 2.8 trillion parameters, around 50 billion are active per token, so less than 2%. It picks 16 of its 896 experts.

What is the difference between open-weight and open source?

With open-weight you get the trained weights to run the model yourself, but not the full recipe and the training data underneath. With true open source you would get that recipe as well.

Can a company use Kimi K3?

Yes, through online services and the API, although that currently runs via Singapore and is not readily suitable for sensitive data. Once the weights are released on 27 July 2026, European parties can host the model under their own rules.

🎧 Want to listen back to the full conversation (in Dutch)? Listen to the BNR fragment here.