原文 · 未翻译
Mistral AI brings the strongest open generative models to the developers, along with efficient ways to deploy and customise them for production.
We're opening a beta access to our first platform services today. We start simple: la plateforme serves three chat endpoints for generating text following textual instructions and an embedding endpoint. Each endpoint has a different performance/price tradeoff.
The two first endpoints, mistral-tiny and mistral-small, currently use our two released open models; the third, mistral-medium, uses a prototype model with higher performances that we are testing in a deployed setting.
We serve instructed versions of our models. We have worked on consolidating the most effective alignment techniques (efficient fine-tuning, direct preference optimisation) to create easy-to-control and pleasant-to-use models. We pre-train models on data extracted from the open Web and perform instruction fine-tuning from annotations.
Mistral-tiny. Our most cost-effective endpoint currently serves Mistral 7B Instruct v0.2, a new minor release of Mistral 7B Instruct. Mistral-tiny only works in English. It obtains 7.6 on MT-Bench. The instructed model can be downloaded here .
Mistral-small. This endpoint currently serves our newest model, Mixtral 8x7B, described in more detail in our blog post . It masters English/French/Italian/German/Spanish and code and obtains 8.3 on MT-Bench.
Mistral-medium. Our highest-quality endpoint currently serves a prototype model, that is currently among the top serviced models available based on standard benchmarks. It masters English/French/Italian/German/Spanish and code and obtains a score of 8.6 on MT-Bench. The following table compare the performance of the base models of Mistral-medium, Mistral-small and the endpoint of a competitor.
Mistral-embed, our embedding endpoint, serves an embedding model with a 1024 embedding dimension. Our embedding model has been designed with retrieval capabilities in mind. It achieves a retrieval score of 55.26 on MTEB.