Join our daily and weekly newsletters for the latest updates and exclusive content on industry-leading AI. learn more
Meta Platforms has created a smaller version of its Llama artificial intelligence model that can run on smartphones and tablets, opening up new possibilities for AI beyond the data center.
The company today announced compressed versions of its Llama 3.2 1B and 3B models. It runs up to 4 times faster while using less than half the memory of previous versions. Based on Meta testing, these smaller models perform almost as well as larger models.
This advancement uses a compression technique called quantization to simplify the mathematical calculations that power AI models. Meta combines two methods: Quantization-Aware Training (QLoRA) using LoRA adapters to maintain accuracy and SpinQuant to improve portability.
This technological achievement solves the important problem of running advanced AI without requiring large-scale computing power. Until now, advanced AI models required data centers and specialized hardware.
Testing on a OnePlus 12 Android smartphone showed that the compressed model is 56% smaller, uses 41% less memory, and can process text more than twice as fast. This model can handle up to 8,000 characters of text, which is sufficient for most mobile apps.
Meta’s compressed AI models (SpinQuant and QLoRA) show significant speed and efficiency improvements compared to standard versions when tested on Android smartphones. Smaller models run up to 4x faster while using half the memory. (Credit: Meta)
Big tech companies race to define the future of AI mobile
The release of Meta intensifies a strategic battle between tech giants to control how AI works on mobile devices. While Google and Apple have taken a measured and controlled approach to mobile AI, maintaining tight integration with their operating systems, Meta’s strategy is markedly different.
By open sourcing these compression models and partnering with chipmakers Qualcomm and MediaTek, Meta bypasses traditional platform gatekeepers. Developers can build AI applications without waiting for Google’s Android updates or Apple’s iOS features. This move reflects the early days of mobile apps, where open platforms dramatically accelerated innovation.
Our partnerships with Qualcomm and MediaTek are particularly important. These companies power most of the world’s Android smartphones, including devices in emerging markets where Meta sees growth potential. By optimizing the model for these widely used processors, Meta allows its AI to run efficiently not only on premium devices, but also on mobile phones across a wide range of price points.
The decision to distribute through both Meta’s Llama website and Hugging Face, an increasingly influential AI model hub, demonstrates Meta’s commitment to reaching developers where they are already working. . This dual-distribution strategy could help Meta’s compression model become the de facto standard for mobile AI development, in the same way that TensorFlow and PyTorch have become the standard for machine learning.
The future of AI in your pocket
Today’s Meta announcement points to a larger shift in artificial intelligence: a shift from centralized computing to personal computing. Cloud-based AI will continue to handle complex tasks, but these new models hint at a future where mobile phones can process sensitive information privately and quickly.
Timing is important. Technology companies are facing increasing pressure around data collection and AI transparency. Meta’s approach addresses both concerns by making these tools open and running directly on the phone. Your phone will soon be able to handle tasks like document summarization, text analysis, and creative writing, instead of a distant server.
This reflects other important changes in computing. Just as processing power has moved from mainframes to personal computers and computing has moved from desktops to smartphones, AI seems poised to move to personal devices. Meta’s bet is that developers embrace this change and create applications that blend the convenience of mobile apps with the intelligence of AI.
Success is not guaranteed. These models still require a powerful smartphone to work properly. Developers must weigh the privacy benefits against the inherent power of cloud computing. And Meta’s competitors, particularly Apple and Google, have their own visions for the future of AI in mobile phones.
But one thing is clear. That means AI is being released from the data center, phone by phone.
VB Daily
Be in the know! Get the latest news in your inbox every day
Thank you for subscribing. Check out other VB newsletters here.
An error has occurred.
