This sentiment analyser uses the Cardiff NLP XLM-T model, an XLM-R system trained on roughly 198 million multilingual tweets and fine-tuned for sentiment in Arabic, English, French, German, Hindi, Italian, Portuguese, and Spanish. The output labels mean unfavorable overall tone, no clear positive or negative lean, and favorable overall tone. They do not identify specific emotions, truth, urgency, or the writer's intent.
The first analysis downloads quantized ONNX model weights plus tokenizer files, about 297 MB in total. A browser worker keeps inference off the main UI thread. WebGPU is preferred when available; the WebAssembly CPU backend is the slower fallback. Browsers normally cache the files, but private browsing, storage cleanup, or low disk space can trigger another download.