Gemma 3n
Gemma 3n is our state-of-the-art open multimodal model, engineered for on-device performance and efficiency. Made for responsive, low-footprint local inference, Gemma 3n empowers a new wave of intelligent, on-the-go applications. It analyzes and responds to combined images and text, with video and audio coming soon. Build intelligent, interactive features that put user privacy first and work reliably offline. Mobile-first architecture, with a significantly reduced memory footprint. Co-designed by Google's mobile hardware teams and industry leaders. 4B active memory footprint with the ability to create submodels for quality-latency tradeoffs. Gemma 3n is our first open model built on this groundbreaking, shared architecture, allowing developers to begin experimenting with this technology today in an early preview.
Learn more
LiteRT
LiteRT (Lite Runtime), formerly known as TensorFlow Lite, is Google's high-performance runtime for on-device AI. It enables developers to deploy machine learning models across various platforms and microcontrollers. LiteRT supports models from TensorFlow, PyTorch, and JAX, converting them into the efficient FlatBuffers format (.tflite) for optimized on-device inference. Key features include low latency, enhanced privacy by processing data locally, reduced model and binary sizes, and efficient power consumption. The runtime offers SDKs in multiple languages such as Java/Kotlin, Swift, Objective-C, C++, and Python, facilitating integration into diverse applications. Hardware acceleration is achieved through delegates like GPU and iOS Core ML, improving performance on supported devices. LiteRT Next, currently in alpha, introduces a new set of APIs that streamline on-device hardware acceleration.
Learn more
Paraspeech
Paraspeech is a speech-to-text app for Mac and iOS that turns spoken thoughts into clean text with a simple hold, speak, and release workflow. On Mac, users press and hold a hotkey in the app where they want to write, speak naturally, and release; Paraspeech processes the recording and attempts to insert the result directly into the active editable field, with clipboard fallback for unsupported fields. Apple Silicon Macs can use supported local speech modes, allowing audio to be transcribed on-device and work offline after setup, while eligible cloud-backed paths are also available depending on the selected backend. Fast local models cover English, Japanese, Mandarin Chinese, and 25-language dictation, while Multilingual Large expands coverage to more than 100 languages where available. AI Rewriting can turn rambling speech into cleaner, formatted text through Cloud Cleanup or, where supported, an on-device rewrite model.
Learn more
Gemini Audio
Gemini Audio is a set of advanced real-time audio models built on Gemini's architecture, designed to enable natural, fluid voice interaction and expressive audio generation through simple language prompts. It supports conversational experiences where users can speak, listen, and interact with AI in a seamless loop, combining understanding, reasoning, and response generation in audio form. It is capable of both analyzing and generating audio, allowing applications such as speech-to-text transcription, translation, speaker identification, emotion detection, and detailed audio content analysis. They are optimized for low-latency, real-time use cases, making them suitable for live assistants, voice agents, and interactive systems that require continuous, multi-turn dialogue. Gemini Audio also integrates advanced capabilities like function calling, enabling the model to trigger external tools and incorporate real-time data into responses.
Learn more
