AI Energy Score v2: Refreshed Leaderboard, now with Reasoning (huggingface.co)

The AI Energy Score project has unveiled a refreshed leaderboard, incorporating a new group of text generation models and introducing reasoning as a benchmarked task. This update highlights the growing importance of evaluating the energy efficiency of AI models, aligning with global initiatives focused on sustainable AI development and policy-making. Having initially launched in February 2025, the leaderboard allows for comparisons across multiple modalities using standardized datasets and the latest GPUs, marking a significant stride towards creating a unified framework in a field where disparate methodologies have hindered progress. The addition of reasoning capabilities underscores a critical energy consideration in AI model performance; reasoning models consume, on average, 30 times more energy than those without. For example, models like DeepSeek-R1 and Phi-4 reveal stark energy discrepancies when activated, utilizing between 150 to 700 times more energy. This finding complicates the assumption that smaller models are inherently more efficient, as their reasoning intensity varies significantly. The AI Energy Score's streamlined benchmarking approach, now open-sourced as AI Energy Benchmarks, will facilitate energy assessment across diverse hardware configurations. Notably, Salesforce has integrated the AI Energy Score into its internal processes, setting a precedent for energy transparency in AI development. As the initiative evolves, the community's engagement is pivotal in fostering responsible AI innovation within sustainable boundaries.