French AI startup Mistral is introducing recent options for customizing AI models, including paid plans that allow developers and firms to optimize their generative models for specific use cases.
The first is self-service. Mistral has released a Software Development Kit (SDK), Mistral Finetuneto fine-tune its models on workstations, servers and small data center nodes.
In the readme file for the SDK's GitHub repository, Mistral points out that the SDK is optimized for multi-GPU setups, but might be scaled all the way down to a single Nvidia A100 or H100 GPU for fine-tuning smaller models just like the Mistral 7B. Fine-tuning a dataset like UltraChat, a group of 1.4 million dialogs using OpenAI's ChatGPT, takes about half an hour using Mistral-Finetune on eight H100s, Mistral says.
For developers and enterprises preferring a more managed solution, there’s Mistral's newly launched fine-tuning services, available through the corporate's API. They are currently compatible with two Mistral models, Mistral Small and the aforementioned Mistral 7B. Mistral says that fine-tuning services will gain support for more models in the approaching weeks.
Finally, Mistral is introducing bespoke training services, currently available only to pick out customers, to optimise each Mistral model for a corporation’s apps using its data. “This approach enables the creation of highly specialised and optimised models for his or her specific domain,” the corporate explains in a post on its official Blog.
Mistral, which recently raised around $600 million at a $6 billion valuation, in keeping with my colleague Ingrid Lunden, is backed by investors including DST, General Catalyst and Lightspeed Venture Partners, and is little question seeking to increase revenue as the corporate faces significant—and increasing—competition within the generative AI space.
Since Mistral unveiled its first generative model in September 2023, several more have been released, including a Code generating modeland has introduced paid APIs. However, the corporate has not disclosed what number of users it has or how much revenue it generates.