Mistral AI Models

Mistral AI develops fully open source foundation models. They also expose their models behind a hosted API including ones not released as OSS.

Supported Variants

Airtrain supports 5 variants of Mistral AI models.

open-mistral-7b

This is an instruction-tuned variant of Mistral 7b. It uses open datasets available on Hugging Face. To see the full model card, refer to Hugging Face.

open-mixtral-8x7b

This is an instruction-tuned variant of Mixtral-8x7b, a sparse mixture of experts model with performance rivaling GPT-3.5-turbo. To see the full model card, refer to Hugging Face

open-mixtral-8x22b

This is an instruction-tuned variant of Mixtral-8x22b, a sparse mixture of experts model with performance surpassing GPT-3.5-turbo. To see the full model card, refer to Hugging Face

mistral-small-latest

mistral-small is a model Mistral only offers via their API. It was originally backed by the same model as open-mixtral-8x7b, but has since diverged.

mistral-medium-latest

mistral-medium is a model Mistral only offers via their API. It is rumored to be the un-quantized precursor to the leaked Miqu model.

mistral-large-latest

mistral-large is a model Mistral only offers via their API. It is considered a GPT-4 class model, though it slightly under-performs GPT-4 on most standard benchmarks.