The Mistral fine-tuning API and SDK allow users to fine-tune and deploy custom Mistral models. This new feature provides performance, speed, and fine editorial control for AI applications. Users can adapt Mistral AI models to their specific needs, reducing the cost and expertise required for customizing generative AI models at scale. The fine-tuning can be done on the user's infrastructure or through Mistral's managed fine-tuning services.

To use the open-source fine-tuning SDK for Mistral models, developers can leverage the mistral-finetune codebase. This lightweight and efficient codebase is built on the LoRA training paradigm, which allows for memory-efficient and performant fine-tuning. Developers can fine-tune all open-source Mistral models on their infrastructure without sacrificing performance or memory efficiency.

Mistral's serverless fine-tuning services on la Plateforme offer several benefits:

• No Infrastructure Hassle: These services eliminate the need for users to manage their own infrastructure.
• Cost-Effective: Leveraging Mistral's unique fine-tuning techniques, these services provide fast and cost-effective model adaptation.
• Efficient Deployment: The use of LoRA adapters prevents forgetting base model knowledge and allows for efficient serving.
• Compatibility: These services are currently compatible with Mistral 7B and Mistral Small models, with plans to add more models in the coming weeks.

Mistral's custom training services involve fine-tuning Mistral AI models on a customer's specific applications using their own proprietary data. This approach enables the creation of highly specialized and optimized models for particular domains. Advanced techniques such as continuous pretraining are used to include proprietary knowledge within the model weights themselves. These services are available to a select group of customers and are tailored to meet their specific needs.