The time has come for the customization of major language models. Mistral AI understands this well. The French start-up now offers “fine-tuning” of its models from its platform. The objective is always the same: to push for ultra-personalization to reach a greater number of companies and, more broadly, equally varied sectors.
To adapt to demand, Mistral AI offers three ways to proceed. First, via an open Source SDK: for developers who want to refine Mistral’s open Source models on their infrastructure, they can rely on mistral-finetune, a dedicated codebase built on LoRA. The startup promises powerful fine-tuning capabilities without the need to sacrifice performance or memory efficiency.
For maximum efficiency, it is recommended to use an A100 or H100 GPU. The codebase is optimized for multi-GPU-single-node training setups, but for smaller models, such as the 7B, a single GPU is sufficient, notes Mistral
A more managed or tailor-made service, users have the choice
The second option consists of a more managed solution directly in the Mistral AI platform. These services leverage Mistral’s fine-tuning techniques and are available through the company’s API. The start-up specifies that it uses “LoRA adapters under the hood to avoid losing knowledge of the base model and provide efficient service.” In this case, only two of the company’s models are compatible: Mistral 7B and Mistral Small. Ultimately, the company plans to make others compatible with its services.
Finally, the third possibility is to use tailor-made services. Users can fine-tune Mistral’s AI models on specific applications using their own data, a boon for niche areas for example. Mistral says it offers advanced techniques such as continuous pre-training to include proprietary knowledge in the weights of the models themselves. Please note, however, that only a restricted group of customers will have access to this option.
The choice of personalization to attract customers
Obviously, Mistral AI is not the only company focusing on personalization to increase its market share. Other companies are making this bet, starting with Snowflake – also an investor in the start-up – and OpenAI. At the end of April, the data cloud specialist launched its own LLM called Arctic which users can customize using their favorite frameworks. These include Nvidia NIM with TensorRT-LLM, vLLM and Hugging Face.
Another player – more discreet – has made this personalization its main business. This is the start-up AdaptiveML which raised 20 million euros last March. Based in France and the United States, the company develops language models adapted for each business and continuously improves them with user interactions. She wants to open an R&D center in Paris, but also improve her technology, to vary interactions and allow language models to self-critique.
Selected for you
