Mistral built its reputation on open-source AI, and now it's inviting users to build on top of it

On Tuesday, Mistral launched Forge, a system that enables customers to build AI models using their proprietary data, including codebases, internal documentation, compliance policies, operational processes, and more. As a result, these models are tailored to each organization’s unique applications, which, according to Elisa Salamanca, Head of Product at Mistral AI, holds many benefits.

“We've been seeing a couple of key areas where model customization is actually critical: It can be because you need to train it on your proprietary data that the models out there have never seen on the web or this is not publicly available information; it can be because you need very specific behaviors; it can be because you actually need to train models for the edge,” said Salamanca in an interview with The Deep View.

Forge supports training across several phases of model development, including pre-training, post-training, and reinforcement learning, allowing companies to feed domain-specific information, fine-tune behavior, and align models and agents with internal policies.

However, technically, competitors' APIs already allow companies to create and fine-tune custom models. As a result, the biggest differentiator, according to Salamanca, is that Mistral’s models are open-source.

“AI Is changing a lot of things, but customizing models and making sure that you are able to own that AI is what's going to get you the differentiation that you need,” said Salamanca. “Forge is actually the critical piece there. It's what's going to make you create your own AI that's not the one that your competitor is going to be able to create and that you are going to own it and not rely on another vendor.”

Our Deeper View

Model deprecation is a real issue that should be considered when building AI fine-tuned AI models. Oftentimes, companies will sunset a model because it is too costly to maintain, especially if a newer, more capable one has been released. However, this fails to account for the fact that, while a new model may be better, it will have nuances that necessitate changes to the established model, as seen in the backlash OpenAI received when retiring GPT-4o. While open-source models have many benefits, this is one of the biggest, and a prime example of the gains that model transparency brings.