Mistral, a French AI startup, is now offering customization options for its AI models, including a paid version, to help developers and businesses tailor the models for specific tasks.
Mistral has launched a software development SDK called Mistral-Finetune, which enables fine-tuning on workstations, servers, and small datacenter nodes. According to Mistral, fine-tuning a model like Mistral 7B on a dataset like UltraChat, which comprises 1.4 million dialogs, takes approximately half an hour using eight H100 GPUs.
For a more streamlined experience, Mistral provides fine-tuning services through its API, currently compatible with Mistral Small and Mistral 7B, with plans to support additional models soon. Furthermore, Mistral is introducing custom training services, allowing select customers to tailor any Mistral model to their specific needs using their own data. This approach enables businesses to create highly specialized models for their unique domains.
As Mistral seeks to expand its revenue amidst increasing competition in the generative AI sector, it is reportedly seeking a $600 million investment at a $6 billion valuation from investors like DST, General Catalyst, and Lightspeed Venture Partners. Despite releasing multiple models, including a code-generating model, and introducing paid APIs since its first generative model in September 2023, Mistral has not disclosed its user base or revenue figures.