Mistral AI La Plateforme Release

Mistral AI has launched La Plateforme, a beta API service that provides developers with access to three generative chat endpoints and one embedding model. This release enables efficient deployment and customization of Mistral AI's open generative models for production environments.

Generative Model Endpoints

Mistral AI provides three chat endpoints for generating text based on textual instructions, each offering a different balance of performance and cost:

Mistral-tiny

Mistral-tiny is the most cost-effective endpoint. It currently serves Mistral 7B Instruct v0.2, a minor release of the Mistral 7B Instruct model. This endpoint is limited to English and achieves a score of 7.6 on MT-Bench.

Mistral-small

Mistral-small serves the Mixtral 8x7B model. This model supports English, French, Italian, German, and Spanish, as well as code. It achieves a score of 8.3 on MT-Bench.

Mistral-medium

Mistral-medium is the highest-quality endpoint, serving a prototype model that ranks among the top serviced models based on standard benchmarks. Like Mistral-small, it supports English, French, Italian, German, and Spanish, and code. It achieves a score of 8.6 on MT-Bench.

Embedding Capabilities

Mistral AI has introduced the Mistral-embed endpoint, which serves an embedding model specifically designed for retrieval capabilities. The model features a 1024 embedding dimension and achieves a retrieval score of 55.26 on MTEB.

API Specifications and Integration

La Plateforme's API follows the specifications of popular chat interfaces to facilitate ease of transition for developers. To support integration, Mistral AI provides Python and Javascript client libraries. The API allows for the use of system prompts to implement higher levels of moderation on model outputs for specific application requirements.

Technical Implementation and Alignment

Models served on La Plateforme are instructed versions of Mistral AI's base models, which are pre-trained on open Web data and then undergo instruction fine-tuning from annotations. The alignment process utilizes a combination of efficient fine-tuning and direct preference optimization (DPO) to ensure the models are easy to control and pleasant to use.

Availability and Infrastructure

La Plateforme is currently in beta access, with registration open to all users as capacity is progressively ramped up. Mistral AI acknowledged NVIDIA for their support in integrating TensorRT-LLM and Triton, specifically in making a sparse mixture of experts (MoE) compatible with TRT-LLM.