Choose a Small, Efficient Model (web.dev)

Right-size your AI: prototype with a large foundation model to prove feasibility, then deploy the smallest model that meets your success criteria. This approach reduces energy use, latency, cost, and privacy risk (on-device inference), while preserving user experience. The practical workflow is: validate the task with a large model (e.g., Gemini 2.5 Pro or other foundation models), define measurable success criteria with test inputs/ideal outputs, evaluate models from small→large (using prompt engineering and larger-model “ground truth” comparisons), and pick the smallest model that produces acceptable results. The guidance also maps common tasks to lightweight model families so teams can pick right-sized architectures: MobileNet/ResNeXt for image classification, YOLOv8/DETR for object detection, MoveNet/BlazePose for pose, MediaPipe/OpenPose for face/hand keypoints, SAM/Mask R-CNN for segmentation, diffusion models like Nano Banana for image generation, Wav2Vec2/AudioMAE for audio classification, Whisper/Kyutai for STT, BERT/DistilBERT for text classification, and compact LMs (Gemma Nano, Granite 1.5B, Qwen 3.4B) for translation or on-device conversational tasks. The piece stresses transparency—model cards, IBM factsheets, Hugging Face’s carbon metrics and AI Energy Score—and urges customers to demand disclosure of training/inference costs to push adoption of sustainable, performant AI.