Edge0

Edge0: On-device speech and language models

Edge0 builds speech and language models designed to run on consumer hardware, from phones and laptops down to earbuds, glasses and toys. Its Audio8 family covers speech recognition and zero-shot text-to-speech, and its Edge0 language models stream expert weights from SSD so a 35B model runs in about 3 GB of memory.

Key Features

• Audio8 ASR family: Three speech recognition models: 3B for accuracy on noisy, far-field and long recordings, 0.6B for balanced real-time and batch transcription, and 0.1B for local recognition on phones and edge devices.
• Multilingual recognition: The 3B and 0.6B models cover 19 languages and handle mixed-language speech, such as Chinese and English in the same sentence, without switching models.
• On-device ASR builds: Audio8 ASR 0.1B ships as ONNX Runtime and Apple Neural Engine builds with INT8 and INT4 quantization, so audio never leaves the device.
• Zero-shot voice cloning: Audio8 TTS 0.6B generates speech in a new voice from one short reference recording and its transcript, with no per-speaker fine-tuning.
• Cross-lingual TTS: The 0.6B TTS covers 11 languages, including English, Chinese, Cantonese, French, German, Japanese and Spanish, and keeps the same voice when the text language changes.
• Compact TTS for CPUs: Audio8 TTS 0.1B is a 170M-parameter model with a complete ONNX INT8 inference path for local CPUs and offline use.
• OpenAI-compatible serving: The 0.6B TTS is built for GPU serving through SGLang Omni, with dynamic batching and an OpenAI-compatible speech endpoint.
• On-device language models: Edge0-35B and Edge0-8B are sparse mixture-of-experts models that load experts from SSD on demand and run through MLX on Apple Silicon.
• Open weights: Models are published on Hugging Face and GitHub, most under the Apache 2.0 license; Audio8 ASR 0.1B uses a non-commercial CC BY-NC 4.0 license.

Use Cases

• Wearables and earbuds: Voice commands handled on the device, with complex requests sent to the cloud only when needed.
• Robots and smart toys: Offline voice interaction for embodied AI, robot chips and connected toys.
• Transcription pipelines: Meeting and interview transcription, subtitles, contact-center QA and real-time captions.
• Private mobile apps: Voice input, field notes and offline captions where audio must stay on the phone.

Getting Started

• Website: https://edge0.ai/
• Models: https://edge0.ai/models
• Hugging Face: https://huggingface.co/Edge0
• GitHub: https://github.com/Edge0-AI

Edge0 suits device makers and developers who want speech recognition, voice generation and language models to run locally, offline and without per-use cloud costs.