Mixture of experts (MoE) is a way of building AI models where the model is split into many smaller specialist parts, called experts, and only a few of them switch on for each word it writes. That lets a model hold a huge amount of knowledge while doing far less work per answer, which is why most of today’s largest open models, from DeepSeek to Kimi, use it.
How does mixture of experts work, in simple terms?
Picture a hospital. It employs hundreds of specialists, but each patient sees only the two or three they need, chosen by a receptionist at the front desk. An MoE model works the same way. A small component called the router looks at each token and sends it to the handful of experts best suited to it. The rest stay idle.
What do “total” and “active” parameters mean?
Model sizes for MoE models come as two numbers. Total parameters are everything the model contains. Active parameters are the part used for each token. For example, GLM-5.2 has 744 billion parameters in total but uses only about 40 billion per token, according to the colibri project. That is why a model can be enormous on disk yet respond almost as fast as a much smaller one.
Why do AI companies use MoE?
- Cheaper to run: less computing per answer means lower costs, one reason DeepSeek trains and serves models so cheaply
- Cheaper to train: each training step only updates part of the model
- More knowledge: the total size can grow without making every answer slower
- Specialisation: different experts can get better at different kinds of text, such as code or maths
What are the downsides of MoE models?
Memory. Even though only a few experts are used at once, all of them usually have to sit somewhere the computer can reach quickly, so MoE models need a lot of memory. Training is also trickier: if the router keeps choosing the same favourite experts, others never learn. Engineers add balancing rules to spread the work.
Clever tools now work around the memory problem. Some keep the active experts in RAM and stream the rest from a fast SSD, which is how people run trillion-parameter models on a laptop, as shown in our colibri guide.
Which AI models use mixture of experts?
Mixtral from Mistral popularised the idea in open models in late 2023. Since then MoE has become standard for large open-weight releases, including DeepSeek’s V4 family, Moonshot’s Kimi K3, Zhipu’s GLM-5 series, Qwen and Mistral Large 4. Many closed models are widely believed to use it too, though their makers rarely confirm architecture details.
Is an MoE model better than a regular model?
Not automatically. A dense model, where every part works on every token, can be simpler to run on small hardware and easier to fine-tune. MoE shines at large scale, where its efficiency savings matter most. When comparing models, look at both total and active parameters, then check real benchmarks, as explained in small vs large AI models. For more terms like this, see our AI glossary.




