Mistral, the Tailor That Builds Your Own Model

Mistral AI has introduced model customization features to la Plateforme, enabling users to gain control over performance, speed, and editing. This allows models to be optimized for specific domains, achieving large-model-level performance while improving deployment cost and speed.

Three approaches are offered for model specialization:

• mistral-finetune SDK: A lightweight codebase for developers who want to train open-source models on their own infrastructure, using the memory-efficient LoRA training method.
• Serverless fine-tuning service: A managed service available on la Plateforme without the burden of infrastructure management. It is compatible with Mistral 7B and Mistral Small, and uses LoRA adapters to enable efficient serving while preserving the model's existing knowledge.
• Custom training service: A tailored service for enterprise customers, offering advanced techniques such as continuous pretraining, which directly embeds knowledge into the model weights using the customer's own data.

Meanwhile, the Mistral Fine-Tuning Hackathon, where participants can experiment with the new fine-tuning API, will run from June 5 to 30, 2024.