French AI startup Mistral is introducing new AI model customization options, including paid plans, to allow developers – and businesses – to fine-tune its generative models for particular use cases.
The first is self-service. Mistral has released a software development kit (SDK), Fine adjustment of the Mistralto refine its models on workstations, servers and small nodes in data centers.
In the SDK’s GitHub repository readme, Mistral notes that the SDK is optimized for multi-GPU setups but can be scaled down to a single Nvidia A100 or H100 GPU to fine-tune smaller models like the Mistral 7B. Fine-tuning a dataset like UltraChat, a collection of 1.4 million dialogs with OpenAI’s ChatGPT, takes about half an hour using Mistral-Finetune on eight H100s, Mistral says.
For developers and businesses that prefer a more managed solution, Mistral’s recently launched fine-tuning services are available through the company’s API. Compatible with two of Mistral’s models for now, Mistral Small and the aforementioned Mistral 7B, Mistral says that fine-tuning services will be supported by more of its models in the coming weeks.
Finally, Mistral is launching custom training services – currently only available to select customers – to refine any Mistral model for an organization’s applications using their data. “This approach allows the creation of highly specialized models optimized for their specific domain,” explains the company in a post on its official website. Blog.
Mistral, which my colleague Ingrid Lunden recently reported looking to raise around $600 million at a valuation of $6 billion from investors including DST, General Catalyst and Lightspeed Venture Partners, is undoubtedly looking to increase revenue as it faces considerable competition – and growing – in the field of generative AI.
Since Mistral unveiled its first generative model in September 2023, it has launched several others, including one code generation model, and deployed paid APIs. But it hasn’t revealed how many users it has, or what its revenue looks like.
