Open Source JavaScript AI Models

Kokoro-82M is an open-weight, lightweight text-to-speech (TTS) model featuring 82 million parameters, developed to deliver high-quality voice synthesis with exceptional efficiency. Despite its compact size, Kokoro rivals the output quality of much larger models while remaining significantly faster and cheaper to run. Built on StyleTTS2 and ISTFTNet architectures, it uses a decoder-only setup without diffusion, enabling rapid audio generation with low computational overhead. Kokoro supports multiple voices and languages and is compatible with environments like Google Colab or production APIs. It was trained on a few hundred hours of permissively licensed and synthetic audio paired with IPA phoneme labels, ensuring broad legal usability. Licensed under Apache 2.0, Kokoro is deployable in commercial, research, and personal projects, including those with monetized outputs under $1M in annual revenue.

Nanonets-OCR-s is an advanced image-to-markdown OCR model that transforms documents into structured and semantically rich markdown. It goes beyond basic text extraction by intelligently recognizing content types and applying meaningful tags, making the output ideal for Large Language Models (LLMs) and automated workflows. The model expertly converts mathematical equations into LaTeX syntax, distinguishing between inline and display modes for accuracy. It also generates descriptive <img> tags for images like logos, charts, and graphs, enabling better interpretation by downstream systems. Signatures and watermarks are detected and isolated within dedicated tags to maintain document integrity, which is vital for legal and business uses. Form elements like checkboxes and radio buttons are converted into standardized Unicode symbols for consistent handling. Additionally, complex tables are extracted and formatted in both markdown and HTML to support versatile document processing.