French AI innovator Mistral is rolling out advanced customization options for its AI models, now offering both self-service and managed solutions to accommodate developers and enterprises looking to fine-tune these generative models for specific applications.
The flagship addition is Mistral-Finetune, a software development kit (SDK) tailored for optimizing Mistral’s models across various computational setups, including workstations, servers, and small datacenter nodes.
According to the SDK’s GitHub documentation, Mistral-Finetune is optimized for multi-GPU configurations but can also operate efficiently on a single Nvidia A100 or H100 GPU when fine-tuning smaller models like the Mistral 7B. For instance, fine-tuning on a dataset such as UltraChat, which consists of 1.4 million dialogs with OpenAI’s ChatGPT, takes approximately 30 minutes when using Mistral-Finetune in a setup with eight H100 GPUs.
Mistral also introduces managed fine-tuning services via its API, catering to developers and businesses seeking a more streamlined approach. Currently, this service supports Mistral Small and Mistral 7B models, with plans to extend compatibility to more models in the upcoming weeks.
Furthermore, Mistral is offering bespoke training services exclusively for select clients. These services allow organizations to fine-tune any Mistral model using their proprietary data, thereby creating highly specialized and optimized models tailored to their unique domain. In a recent blog post, the company highlighted that this approach facilitates the development of precise and efficient models for specialized applications.
As reported by my colleague Ingrid Lunden, Mistral is currently pursuing a substantial funding round, aiming to secure approximately $600 million at a valuation of $6 billion, with investors like DST, General Catalyst, and Lightspeed Venture Partners showing interest. This funding will likely bolster Mistral’s growth and revenue as the company navigates the competitive landscape of generative AI.
Since the launch of its first generative model in September 2023, Mistral has introduced several models, including a code-generating version, and has begun offering paid APIs. However, the company has yet to disclose user metrics or revenue figures.
