Edge0

Edge0 is an open-source inference framework designed to run large mixture-of-experts language models on devices with limited memory. It streams neural network expert weights from SSD storage instead of loading every parameter into RAM. A trained routing predictor anticipates which experts will be needed and loads their weights ahead of execution. Recover-LoRA helps restore model quality after memory-efficient modifications. The framework supports experimental 8B and 35B-class models and targets consumer computers and mobile devices. Platform-specific inference engines are available for macOS, iOS, Android, and Windows. Its architecture separates model storage from active computation to reduce peak memory usage. Edge0 also provides technical documentation, benchmarks, and experimental model checkpoints.