What Is a Mixture-of-Experts (MoE) AI Model?

Read any AI release note in 2026 and you’ll see numbers like “816B total, 42B active” or “35B-A3B.” These describe mixture-of-experts models, usually shortened to MoE. It’s the design behind many of the largest new models, including several from our October roundup.

You don’t need to understand the maths to make sense of it. Here’s what MoE means, why it’s everywhere, and what it changes for you as a user.

Dense models: everyone works on everything

To understand MoE, start with the alternative. In a dense model, every part of the network is used for every word (more precisely, every token, a chunk of text roughly three-quarters of a word long).

Think of a company where every employee reads every email that comes in. It works, but as the company grows, it gets slow and expensive.

Mixture of experts: send each task to specialists

An MoE model splits much of the network into many smaller sub-networks called experts. A small component called the router looks at each token and sends it to just a few of them, often two to eight out of dozens or hundreds.

Back to the company: now a receptionist reads each email and forwards it to the two or three people best suited to handle it. Everyone else carries on with other work. The company can hire far more specialists without every email costing more. The total headcount grows, but the work done on any single email stays about the same, and that’s the whole trick behind MoE.

A few clarifications:

• “Experts” aren’t neat topics. You won’t find a “history expert” and a “coding expert.” The model learns its own way of dividing work during training, and it’s often hard for humans to interpret.
• Routing happens constantly. Different tokens in the same sentence can go to different experts.
• Some parts stay shared. Typically only certain layers are split into experts. Other parts are used for every token.

Total vs active parameters

This is the key idea behind those release-note numbers.

• Total parameters: everything stored in the model, including every expert.
• Active parameters: the portion actually used for each token.

Some October 2026 examples:

Model	Total parameters	Active parameters
Meta Muse Spark Open	~816B	~42B
Reflection Beam 1.0	~501B	~23.4B
Index-Translate-35B-A3B-preview	35B	3B
Liquid AI LFM2.5-24B-A2B	24B	2B

The naming pattern “35B-A3B” simply means 35 billion total, 3 billion active.

Why it matters:

• Active parameters drive speed and running cost. Each token only goes through the active part, so an 816B MoE model can run closer to the cost of a much smaller dense model.
• Total parameters drive memory. All the experts still have to be loaded somewhere, so you need enough memory to hold the whole model.

Why so many new models use MoE

1. More capability per dollar. A larger total size lets the model store more knowledge and skills, while the active size keeps each request affordable.
2. Cheaper APIs. Lower compute per token helps providers offer low prices, one reason budget models with huge context windows have appeared.
3. Big open models become practical. Open-weight MoE models can be very large yet still run at reasonable speeds on the right hardware.
4. Small MoE models can suit devices. With only a couple of billion active parameters, models like LFM2.5 are designed for on-device AI agents.

The trade-offs

MoE isn’t free:

• Memory hungry. You still need enough memory for all the experts. That’s why running a huge MoE model yourself still requires serious hardware, even if it’s fast once loaded.
• Harder to train. Routers can overload some experts and ignore others, so labs need extra techniques to keep work balanced.
• Fine-tuning can be trickier than with dense models of a similar active size.
• Not automatically better. A strong dense model can outperform a weaker MoE model. The architecture is a means, not a guarantee.

What this means for you

If you use AI through apps or APIs:

• Don’t choose a model just because of its total parameter count. Bigger total size isn’t the same as better results for your tasks.
• Expect MoE to keep pushing prices down, especially for long-context work.

If you want to run models yourself:

• Check total parameters (and the file size) to see if the model fits in your computer’s memory.
• Check active parameters to estimate how fast it will run.
• Start with a small model. Our guide to running an AI model on your laptop walks through it.

For everyone: test candidate models on ten real tasks before switching. Benchmarks published at launch usually come from the company selling the model.

Key takeaways

• A mixture-of-experts model routes each token to a few specialist sub-networks instead of using the whole model.
• Total parameters measure the full size; active parameters measure how much is used per token.
• Active size mostly drives speed and cost; total size mostly drives memory needs.
• MoE makes big, capable models cheaper to run, which is why it’s so common in 2026.
• Judge models on your own tasks, not on parameter counts.

Deciding between hosted and self-run models? Read our comparison of open-weight vs closed AI models, and see what a context window is for the other number that fills release notes.