Mistral AI Discloses Carbon Footprint of AI Model Training

“We cannot ignore the climate issues of an industry that everyone knows has an environmental impact.” This is how Audrey Herblin-Stoop, Vice President of Mistral AI, justifies the publication on Tuesday, July 22nd, of the first report on CO2 emissions and the consumption of water and materials from the training of one of its large language processing models.

NGOs and researchers, such as Sasha Luccioni from the start-up Hugging Face, have complained about the lack of information from artificial intelligence companies regarding the ecological impact of their activities. Mistral AI’s study is unique in that it involved external reference partners, the environmental agency Ademe and the specialized firm Carbone 4. With this initial study, the company is advocating for greater “transparency” and “comparability” among players in the sector.

The study first evaluates the training of Mistral Large 2, “chosen because it was the heaviest of our AI models,” explains Ms. Herblin-Stoop; 86% of the carbon impact is related to the electricity required for this phase. The balance is estimated at 20.4 kilotonnes of CO2 equivalent (ktCO2e). This is roughly equivalent to 3,375 airplane trips around the world for one passenger or 95 million kilometers by car, according to Ademe’s calculator. “The figures are high, that’s a fact, but the sector needs an objective assessment in order to then take action,” emphasizes Ms. Herblin-Stoop. Carbone 4 notes that lighter models have a reduced impact “proportional to their size,” by a spectacular factor of 100 for the company’s smallest model, Ministral, which remains effective for simple requests.

---

Enjoyed this post by Thibault Helle? Subscribe for more insights and updates straight from the source.