What are parameters in an AI model?

What a parameter is

Google's machine learning glossary defines parameters as the weights and biases that a model learns during training. They differ from hyperparameters. Those are values that a person or a tuning service supplies to the model, such as the learning rate.

Once training ends, the values are fixed. The Open Source Initiative says these final weights and biases determine how a model interprets input and generates output. When a developer shares them, others can adapt the model or run it themselves. Such a release is called open weights.

Why the count matters

A model's capacity is the complexity of the problems it can learn. Google's glossary says capacity typically increases with the number of parameters. A 2017 paper by Noam Shazeer and six co-authors makes a similar point. It says a neural network's capacity to absorb information is limited by its number of parameters.

Counts vary widely. Mistral's Mistral 3 release of December 2025 ranged from 3 billion parameters to 675 billion. Mistral aimed the three smallest models at edge and local use. It said a compressed version of the largest could run on a single node with eight A100 or H100 GPUs. GPUs are the chips used to run AI models.

Some newer models pass 1 trillion. Mistral's documentation listed Mistral Large 4 at 1.05 trillion total parameters on October 8, 2026.

What the count does not show

• Not every parameter works at once. In a mixture-of-experts model, only part of the network handles each token, a unit of text. Labs then quote total and active counts. DeepSeek lists its V4-Pro model at 1.6 trillion total and 49 billion active.
• A smaller model can come close. In the same note, DeepSeek describes V4-Flash, a model with 284 billion total parameters. It says that model's reasoning closely approaches V4-Pro. That is the company's own claim.