The emerging French company Mistral offers new options to customize artificial intelligence models, including paid plans, allowing developers and enterprises to fine-tune generative AI models to fit specific use cases.
Precise fine-tuning is crucial for improving the performance of large language models and adapting them to meet the specific needs of the organization.
Fine-tuning leads to achieving accurate and useful model responses, enabling organizations to extract value and accuracy from generative AI applications.
The company has introduced a software development package named Mistral-Finetune to enhance its models on workstations, servers, and modules in small data centers.
Mistral indicates that the software development suite is optimized to work with various settings of graphics processing units, with the ability to scale down to operate on a single A100 or H100 GPU for fine-tuning small models like Mistral 7B.
The company mentions that fine-tuning a dataset like UltraChat accurately takes around half an hour using Mistral-Finetune with eight H100 units.
The new fine-tuning services from Mistral are now available through an API for developers and companies preferring managed solutions.
The fine-tuning services currently support two models from Mistral, namely Mistral Small and Mistral 7B. Mistral states that additional models will be supported for these services in the coming weeks.
Mistral is offering custom training services for the first time, currently available to a limited number of clients, with the goal of adjusting any of its models to meet the needs of enterprise applications using their data.
The company explains that this approach enables them to develop specialized and high-performance models tailored to the enterprise’s domain.
Mistral aims to raise nearly $600 million at an evaluation of up to $6 billion from investors, including DST, General Catalyst, and Lightspeed Venture Partners.
The company seeks to boost its revenues due to intense and growing competition in the field of generative artificial intelligence.
Mistral has launched several additional models since announcing its first generative model in September 2023, including a model for code generation.
It has also introduced paid APIs, although it has not disclosed its user base or revenues.
