AI capabilities have been available on smartphones for a year. Google offers them with Gemini and Apple (to some extent) with Apple Intelligence. However, these functions are limited and handle modest tasks.
Meanwhile, PCs are increasingly running more sophisticated AI models. My recent comparison of DeepSeek R1-14b with models like Llama 3.1-8b or Phi 4-14b showed that these developments run well on a Mac mini M4 with 16 GB of RAM.
With Pixel phones, for example, Google offers its Gemini Nano model in two versions: 1.8B and 3.25B parameters. While decent, they still fall short of the performance of models like the DeepSeek R1-14B and others mentioned.
The issue is that larger models with more parameters (e.g., 14B) require substantial memory. A seven-billion-parameter LLM (7B) typically needs about 8 GB of memory, though having more headroom (e.g., 12 GB) is advisable.
Manufacturers are aware of this. Even Apple has stepped up, with the iPhone 16s moving from 6 GB to 8 GB of RAM, while Google’s Pixel 9 offers up to 16 GB for the same reason. More RAM allows locally executed AI functions to run smoothly.
That memory jump could soon go further. It’s not unreasonable to imagine future phones with at least 32 GB of RAM to support larger AI models, giving users more powerful capabilities.
Of course, memory isn’t the only factor. Phones lack a dedicated GPU to accelerate these tasks, but developers are advancing increasingly powerful NPUs. The combination of these elements signifies a shift toward delivering more versatile local AI models.
In addition to these hardware improvements, potential optimization and “compression” techniques for AI models exist. Quantization—a form of “rounding”—helps reduce the size of LLMs at the expense of some accuracy.
Quantization is popular for using large AI models on smaller devices, lowering hardware requirements and boosting efficiency.
All of this points to a near future where much more powerful AI models reside in our pockets—models that run locally, work without an internet connection, and keep conversations private.
The benefits are compelling. It’s easy to imagine handset manufacturers soon boasting 32 GB phones—or perhaps even more.
Image | Solen Feyissa (Unsplash)
Related | We Made a Complete List of All OpenAI’s Models to Try to Understand Their Names. It’s Hopeless
View 0 comments