Mistral AI's Platform Service, La Plateforme

Mistral AI has opened beta access to La Plateforme, which provides developers with powerful open generative models and efficient deployment methods. This platform offers three chat endpoints for text generation and an embedding endpoint.

The main endpoint configuration is as follows:

• mistral-tiny: The most cost-effective model based on Mistral 7B Instruct v0.2, supporting only English, scoring 7.6 on MT-Bench.
• mistral-small: Using the Mixtral 8x7B model, it supports English, French, Italian, German, Spanish, and code, scoring 8.3 on MT-Bench.
• mistral-medium: A high-performance prototype model currently being tested, showing high performance of 8.6 on MT-Bench along with multilingual and code support.

The embedding service, mistral-embed, is a 1024-dimensional model optimized for Retrieval, scoring 55.26 on MTEB.

The API follows the existing popular chat interface specification, and provides Python and Javascript client libraries. Additionally, moderation of model output can be configured through system prompts. Through collaboration with NVIDIA, TensorRT-LLM and Triton integration has been completed, securing Sparse Mixture of Experts (MoE) compatibility.