Mistral AI Discloses Environmental Impact of AI Model Training

Mistral AI has released a comprehensive report detailing the environmental impact of its large language models, announced on their website. The report, conducted in collaboration with France's environment agency ADEME and consulting firm Carbone 4, quantifies the greenhouse gas emissions from Mistral's flagship models. Training the Mistral Large model emitted approximately 2,200 tons of CO2 equivalent, comparable to the annual emissions of 500 French households.

The study also highlights the ongoing emissions from AI inference, which can vary based on hardware efficiency. Mistral AI advocates for industry transparency and pledges to update these environmental impact reports regularly. This initiative aims to empower users and enterprises to choose more sustainable AI solutions.

The report follows the Frugal AI methodology developed by French standards body AFNOR, emphasizing efficiency and minimal resource use in AI development. Mistral AI calls on competitors like OpenAI and Google to publish their own carbon footprints to enable fair comparisons. The data from this study will be integrated into ADEME's Base Empreinte database, potentially setting a benchmark for AI environmental reporting worldwide.