jina-ai/clip-as-service

Provides runtime flexibility with TensorRT, ONNX, and PyTorch backends for different performance-latency tradeoffs, supporting up to 800 QPS. Implements duplex streaming with non-blocking I/O and automatic load balancing across multiple CLIP models on a single GPU. Integrates natively with Jina and DocArray ecosystems via gRPC, HTTP, and WebSocket protocols, enabling rapid deployment of cross-modal neural search applications.

12,825 stars. No commits in the last 6 months. Available on PyPI.