# Mobile Devices

On-device Sora is an open-source project aimed at enabling efficient video generation on mobile devices (such as the iPhone 15 Pro) using technologies like Linear Proportional Jump (LPL), Time Dimension Tag Merging (TDTM), and Concurrent Inference with Dynamic Loading (CI-DL). Developed from the Open-Sora model, it generates high-quality videos based on text input. Its main advantages include efficiency, low power consumption, and optimization for mobile devices. This technology is applicable in scenarios that require rapid video creation on mobile devices, such as short video production and advertising. The project is currently open source and can be used free of charge.

MobileLLM is an optimized small language model designed for mobile devices, focusing on the creation of high-quality LLMs with less than one billion parameters to cater to practical mobile deployments. Contrary to traditional beliefs, this research emphasizes the importance of model architecture in small LLMs. Through deep and thin architecture, combined with embedding sharing and grouped query attention mechanisms, MobileLLM achieves significant accuracy improvements and introduces a block-level weight sharing method that does not increase model size nor incur high latency costs. Furthermore, the MobileLLM family demonstrates remarkable improvements in chat benchmarks compared to previous small models, and approaches the accuracy of LLaMA-v2 7B in API call tasks, showcasing the potential of small models in common device use cases.

The Llama model is a large language model developed by Meta. Through quantization technology, it reduces model size, increases speed, and maintains quality and security. These models are especially suitable for mobile devices and edge deployments, enabling fast on-device inference on resource-constrained devices while minimizing memory usage. The development of the Quantized Llama model marks an important advancement in mobile AI, allowing more developers to build and deploy high-quality AI applications without requiring extensive computational resources.