Mistral AI now offers three routes for model fine-tuning

Mistral AI has introduced model customization options to its La Plateforme, allowing developers to fine-tune their models to achieve enhanced speed accuracy, performance, and editorial control in their AI applications. There are three options to model fine-tuning: an open-source SDK for fine-tuning on their infrastructure, serverless fine-tuning services on la Plateforme leveraging Mistral's proprietary techniques, and custom training services for enterprise clients.

With the mistral-finetune codebase, developers can fine-tune their open-source models on their infrastructure using the LoRA training paradigm, which results in memory-efficient, performant fine-tuned custom models. In contrast, the serverless option is aimed at those who do not have access to the appropriate infrastructure, or would rather take advantage of Mistral's fine-tuning services, optimized for speed and cost-effectiveness. Currently, both fine-tuning methods support Mistral 7B and Mistral Small, with new models being added soon. The Mistral Fine-Tuning Hackathon, taking place from June 5–30, 2024, will encourage experimentation using the novel fine-tuning API.

Mistral's custom training services are available to selected customers and consist of fine-tuning Mistral AI models on the customer's applications using their proprietary data and advanced techniques to deliver highly specialized models optimized to a particular domain. For more information on these services, prospective customers are advised to contact Mistral's sales team.