GemmaEmbed Swift package optimized for Apple Silicon (github.com)

GemmaEmbed has launched a new Swift package optimized for Apple Silicon that efficiently runs Google's EmbeddingGemma 2 encoders for text, image, and audio. This innovative tool operates independently of third-party dependencies, using Metal kernels for in-process execution. Users can directly input official GGUF weight files to generate precise 768-dimensional embeddings closely aligned with llama.cpp's outputs—achieving a remarkable cosine similarity of 0.99999 for text and images, and 0.99998 for audio. This development is significant for the AI/ML community as it streamlines the embedding process, allowing the integration of various data types (text, images, and audio) into a unified model. With a robust architecture—comprising 24 layers and 271 million parameters—GemmaEmbed leverages advanced features like customizable token limits and efficient batch processing, accommodating up to 8192 tokens per text input. By enabling seamless multimodal analysis and providing compatibility with macOS and Swift 6, this package not only enhances the processing speed on Apple devices but also advances the accessibility and interoperability of machine learning applications across different media types.