TinyCLIP ViT-8M/16

TinyCLIP-ViT-8M-16-Text-3M-YFCC15M is a zero-shot image classification model that performs cross-modal retrieval between images and text without task-specific fine-tuning. It is a distilled CLIP model using affinity mimicking and weight inheritance, designed to balance accuracy and computational efficiency. The model achieves 41.1% ImageNet zero-shot top-1 accuracy while requiring only 2.0 MACs (G) and supporting 4,150 image-text pairs per second throughput. It surpasses the original CLIP ViT-B/16 by 3.5% in ImageNet zero-shot accuracy while using only 8.9% of the parameters. The distillation with weight inheritance also speeds up training by 1.4–7.8× compared to training from scratch.

Key Strengths

• Extreme parameter efficiency: 8M vision encoder and 3M text encoder
• High throughput: 4,150 pairs/s for rapid inference
• Competitive zero-shot accuracy relative to model size

Benchmark Results

Model Variant	ImageNet Acc@1 (%)	MACs (G)	Throughput (pairs/s)
TinyCLIP ViT-8M/16 Text-3M	41.1	2.0	4,150

This model is hosted by gigarouter as a managed, OpenAI-compatible API. No local installation or model loading is required; simply call the API endpoint for zero-shot image classification or image-text similarity tasks.