The French artificial intelligence firm Mistral AI has released a new set of tools aimed at enabling developers to fine-tune AI models effectively. The centerpiece of this update is the Mistral-Finetune SDK, designed to allow the fine-tuning of AI models across various computing environments, such as individual workstations and small data centers.
According to Mistral's detailed GitHub documentation, the SDK is specifically optimized for multi-GPU setups but can also operate with a single Nvidia A100 or H100 GPU for smaller models like Mistral 7B. For example, tuning a dataset like UltraChat, which contains 1.4 million dialogues with OpenAI's ChatGPT, can be done in approximately 30 minutes using eight H100 GPUs.
Managed Fine-Tuning Services
In tandem with the SDK, Mistral has introduced managed fine-tuning services through its API. Initially available for the Mistral Small and Mistral 7B models, the company plans to broaden support to additional models. This managed service allows organizations to fine-tune AI models without the need to handle the infrastructure, streamlining the process.
Custom training services are also being offered, currently restricted to select clients. These services allow businesses to fine-tune any Mistral model using their data, resulting in models that are finely tuned for specific tasks or industries.
Competitive Maneuvers in AI
These new tools and services form part of Mistral's strategy to bolster its position in the competitive generative AI market. Speculation suggests the startup is in the process of raising around $600 million at a valuation of $6 billion from investors like DST, General Catalyst, and Lightspeed Venture Partners. This capital is expected to help Mistral expand its customer base and diversify revenue.
Microsoft has recently entered into a multi-year partnership with the France-based startup Mistral AI, aimed at enhancing the AI capabilities available on its Azure platform. This collaboration marks a potentially important step for Microsoft as it seeks to expand its AI offerings, allowing Azure AI customers to access large language models developed by Mistral AI. The partnership commenced with the integration of Mistral's Large Language Model (LLM) into Azure AI services, and now, the Mistral Small LLM has been made accessible to customers as well.
Since launching its initial generative model in September 2023, Mistral has developed several other models, including one for code generation, and has rolled out paid APIs. However, the company has not yet released any data on user numbers or revenue.
Reducing AI Customization Barriers
The open-source SDK, Mistral-Finetune, employs the LoRA training methodology, designed for optimal memory utilization and performance. The serverless fine-tuning services on their AI development platform, la Plateforme, use proprietary techniques and LoRA adapters to preserve the base knowledge of models while streamlining deployment. For select clients, custom training services enable high-specialization models tailored to unique domains utilizing advanced techniques like continuous pretraining.
