fastText

fastText comes from Meta Research and handles two jobs with unusual efficiency: learning word representations and classifying text. Its word vectors incorporate character n-gram information, so embeddings can be built for out-of-vocabulary words by composing subword units, and pretrained vectors are published for 157 languages trained on Wikipedia and Common Crawl. The supervised classifier trains on billions of words in minutes on ordinary CPUs and performs competitively with deep learning approaches on tasks like sentiment analysis and language identification, while built-in quantization shrinks trained models to a fraction of their size for deployment on constrained devices. The core is implemented in C++11 with a command-line tool and official Python bindings. Released under the MIT license, fastText remains a common choice for practitioners who need a fast, dependable baseline for text classification, language detection, or multilingual embeddings without GPU infrastructure.