Mixtral 8x22B Matches GPT-4 On Your GPU. It Is Open Source. The Frontier Just Moved To Your Basement.

Mistral released Mixtral 8x22B, a 141-billion-parameter mixture-of-experts model, under the Apache 2.0 license. It fits on a single consumer NVIDIA 4090 GPU and matches GPT-4 quality on many tasks. You can download it, run it locally, and never pay an API bill again.

This proves the principle of democratized frontier intelligence through architectural efficiency. The mechanism is mixture-of-experts, or MoE. Instead of activating all 141 billion parameters for every token, the model routes each request through only the relevant subset of expert networks. You get GPT-4 class reasoning with a fraction of the compute cost. The mental model: model quality is no longer locked to model size. Smart architecture beats brute force. The frontier is not a wall. It is a gradient, and open source is sliding down it toward zero.

Mistral, the French AI lab, built and released this model under Apache 2.0. That license means you can use it commercially with no restrictions. The model is available now for anyone with a 4090 and the patience to configure inference software.