Mixtral 8x22B Instruct

Mixtral 8x22B Instruct – Sparse Mixture-of-Experts Instruction-Tuned Model

Mixtral 8x22B Instruct is the instruction fine-tuned version of Mixtral-8x22B-v0.1, a sparse Mixture-of-Experts (SMoE) model released by Mistral AI under Apache 2.0. Its defining architectural property is selective parameter activation: of 141B total parameters across 8 expert networks, only 39B activate per token during inference — roughly 28% of total capacity — enabling throughput characteristics faster than dense models of comparable active size.

Architecture Detail

The SMoE routing mechanism selects which expert sub-networks process each token independently, leaving the remaining experts inactive for that token. This is distinct from dense models where all parameters participate in every forward pass. Mistral officially positions this as delivering performance-per-active-parameter efficiency that outpaces dense open-weight alternatives at the 70B class.

Deployment & Status

Released April 2024 under Apache 2.0 with both base (Mixtral-8x22B-v0.1) and instruct (Mixtral-8x22B-Instruct-v0.1) checkpoints. The API endpoint open-mixtral-8x22b reached retirement on March 30, 2025, with Mistral Small 3.2 designated as the official replacement per Mistral's model documentation.