Mixtral 8x22B latest-model benchmarks

>"The Mixtral 8x22b is a state-of-the-art language model that leverages a cutting-edge architecture called Mixture of Experts (MoE). In this setup, the model comprises eight individual "expert" models, each with 7 billion parameters."

Each with... how many parameters?

>"Key Features: 8 expert models, each with 7 billion parameters (totaling 56 billion parameters)"

8x7?? not 8x22??

>"For instance, on the challenging ARC reasoning task, the Mixtral 8x22b achieved an impressive score of 70.5, surpassing even GPT-4's reported score of 68."

Not a single score on any benchmark shown from GPT4 is 68. The ARC-C score reported is 96.3. What!

Am I going crazy, or is this entire article GPT slop?

*Benchmarks done by the community.

how many shots?