Apple is working to replace the large Gemini model in the iPhone to power the new Siri

It’s impossible to completely avoid AI output when interacting with technology, but Apple has a few. It’s not entirely by choice, though. The iPhone maker has delayed AI-enhanced Siri several times since it was promised in 2024, but partnership with Google will include a well-known partner with Gemini later this year. As we approach International Developer ConferenceApple has been working to bring great AI intelligence to the humble smartphone space. Apple fans may not like the results, however.

Apple has long lamented the need for privacy to run your own AIbut a new report suggests that despite Apple’s best efforts, iPhone’s Gemini makeover will lean heavily on Google and Nvidia in the cloud. The Information reports that Apple’s Gemini-infused Siri will run both on-device and in the cloud, which seems to be changing the privacy preferences of local AI.

With every new chip announcement, we hear about how silicon is designed for AI—even if Apple does it with a focus on upgrading the Neural Engine. You might think from the general language that mobile phones are equipped with AI-type functionality, but it is not. In fact, GPUs in most phones can process more AI tokens than AI-focused NPUs. Tools like Apple’s Neural Engine are designed to make efficient and effective use of AI. Even if the phones have fast AIs, they don’t have the RAM to store large amounts of data.

Even the biggest AI models are still agents in the middle, and this makes local AI very difficult. The AI ​​models that run on phones are physically small, with a few billion parts. Compare that to Google’s latest version of Gemini, which has billions of shares, The Information says. AI models on devices are also “multiplied” for efficiency, making them faster but affecting the accuracy of token generation. All of this adds up to AIs that feel less intelligent than their cloud counterparts, and even great cloud models can be dumb at times.

Amazing Gemini, it’s shrinking

Google has Gemini models optimized for mobile devices, which it calls Gemini Nano. However, these are designed to support standard features such as Magic Cue and text abbreviations. Siri, on the other hand, is supposed to be a conversational assistant – you talk to her and she does things. These are different situations that require a different type of model. On Android, Google doesn’t even bother trying to do this locally. Talking to Gemini always goes to the cloud.