Mistral AI has unveiled Voxtral, a powerful new open-weight model designed to excel at multilingual tasks and efficient inference. With 21 billion parameters and a Mixture of Experts (MoE) architecture, Voxtral activates only 2 of 64 experts per forward pass—bringing high performance while controlling compute costs.

Unlike traditional dense models, Voxtral’s MoE structure allows it to scale intelligently by using different subsets of parameters depending on the input. This enables the model to specialize and deliver strong results across languages, a key focus area for Mistral AI.

According to the official announcement, Voxtral was pretrained on 400B tokens and supports efficient inference at under 15 tokens per second per H100 GPU. It supports 20+ languages, demonstrating competitive reasoning abilities especially in multilingual benchmarks.

Mistral emphasized Voxtral’s strong performance on tasks requiring both general reasoning and domain adaptation. The release also includes the training configuration and weights under the Mistral AI License, continuing the company’s mission to open up access to cutting-edge AI research.

The Voxtral model joins Mistral’s growing suite of open-weight models and further strengthens the Paris-based AI lab’s position in the competitive foundation model landscape.