Meta Platforms has unveiled a streamlined version of its Llama model with two products: Llama3.21B and 3B. This breakthrough allows large-scale language models to operate stably on ordinary smartphones and tablets for the first time. The new version reduces file size by 56% and memory requirements by 41%, while maintaining original processing quality and quadrupling processing speed, enabling continuous text processing of up to 8,000 characters.
During tests on Android phones, Meta's compressed AI models (SpinQuant and QLoRA) showed significant improvements in speed and efficiency compared to standard versions. The smaller models ran four times faster with reduced memory usage. Practical tests on the OnePlus12 demonstrated performance comparable to the standard version while enhancing operational efficiency effectively addressing mobile device computing power issues.
Meta has adopted an open collaboration market strategy, partnering deeply with major mobile processor manufacturers like Qualcomm and MediaTek. The new version will be released simultaneously on the Llama official website and Hugging Face platform, providing developers with convenient access channels. This contrasts sharply with industry giants like Google and Apple, which integrate technologies deeply into their operating systems.
This release signifies a shift from centralized server processing to personal terminal processing, offering better user privacy protection and faster response experiences. While challenges remain regarding device performance requirements and developer platform choices, Meta hopes that this breakthrough will drive the entire industry towards greater efficiency and security through open collaboration, paving new paths for future application development on mobile devices.
