The AI ​​raises a huge change in our mobiles. One that will have (at least) 32 GB of RAM

A year ago our mobiles have AI functions. Google offers them with Gemini and Apple (more or less) with Apple Intelligencebut for now these functions are limited and are reduced to somewhat modest tasks.

However, we are seeing how our PCs have access to more striking models. The recent comparison we made of Deepseek R1-14b with models as it calls 3.1-8b or phi 4-14b showed that these developments could really run well in a Mac Mini M4 with 16 GB of RAM.

However, what happens for example in the Pixel is that Google offers its Gemini Nano modelwhich has two versions: one 1.8b and another 3.25b. They are decent models, but they are still clearly below the benefits of models such as Deepseek-R1-14B and others such as those mentioned.

The problem is that these models, especially when we begin to raise the size and number of parameters (14b, for example), They need memory. And enough. A LLM of seven billion parameters (7b) usually need about 8 GB of memory, although here has some more margin of maneuver (for example, 12 GB) is recommended.

The manufacturers know it, and in fact even Apple has made a small effort there. In the iPhone 16 The jump from 6 to 8 GB has been made largely because of this, and Google Pixel 9 They offer up to 16 GB of RAM precisely for the same reason: that gives maneuver margin so that the functions of the executed in local can function fluidly.

But that jump may soon go more. It does not seem unreasonable to think that sooner rather than later let’s see mobiles with at least 32 GB of RAM precisely to be able to execute larger AI models and thus offer users more powerful options in this regard.

Of course, not only does the amount of memory matter. Our mobile phones do not have a dedicated GPU that can accelerate these tasks, but much is being progressed in the development of NPUs increasingly powerful. The combination of both elements seems to make possible an important change in offering local models of increasingly versatile.

These hardware improvements in our mobiles also join possible techniques of optimization and “compression” of AI models. The quantization, a kind of “rounding”, allows large language models (LLM) to see their size reduced, yes, from the loss of a certain level of precision.

The quantization It is an already very popular process when being able to use large models in more modest machines, and in addition to reducing hardware requirements it also allows to gain efficiency.

All this suggests a not too distant future in which we will have much more powerful models in our pocket. Models that we can run at home, which we can even use without internet connection and that will also maintain the entire conversation in private.

There are many interesting advantages. Too many Not to think that we may soon see how mobile manufacturers presume 32 GB mobile. Or who knows if even more.

In Xataka | The new Gemini demonstrates a Google ambition: that we talk without stopping with our mobile