Towhee

Towhee is a free, open-source machine-learning framework for building data processing pipelines that turn unstructured data — images, video, text, audio and more — into AI-ready embeddings. Its motto is “x2vec, Towhee is all you need”: describe your pipeline with a Pythonic API, pick from hundreds of pre-trained models, and Towhee handles optimisation and execution.

Overview

Towhee is an open-source project on GitHub with about 3,500 stars, developed by the creators of Milvus and last updated in 2024. It provides more than 700 pre-trained embedding models across computer vision, NLP, multimodal, audio and medical domains, covering over 140 model architectures such as BERT, CLIP and ViT. Pipelines are written with a method-chaining Python API that feels like working with tabular data, and Towhee applies end-to-end optimisations — from decoding to inference — that can make execution many times faster. It integrates naturally with vector databases like Milvus, making it a practical building block for image retrieval, video analytics and multimodal search systems.

Key Features

• Pythonic, method-chaining API for building ML data pipelines
• 700+ pre-trained embedding models across 5 domains and 15 tasks
• Supports nearly 20 unstructured data modalities, from images to 3D molecules
• End-to-end pipeline optimisation for production-ready performance
• Native fit with Milvus and the wider vector-search ecosystem
• Schema-based processing that makes unstructured data feel tabular

💳 Pricing

Towhee is completely free and open source under the Apache-2.0 license. There are no paid tiers — you run it on your own infrastructure.

Who Is It For

Towhee is for ML engineers and data scientists who build embedding pipelines for search, retrieval and analytics over unstructured data, and want production-grade speed without hand-optimising every step.