Meet the AI-first phones powered by small models

Surprise! AI agents are now on your mobile phones, too. The latest “AI Phone,” a joint product from T-Mobile parent company Deutsche Telekom and AI giant Perplexity, took center stage at this year’s Mobile World Congress.

But while Google and Apple have been adding AI features to phones for years, the AI-first phones launching in 2025 may signal a bigger shift, say IBM experts like Kaoutar El Maghraoui, a Principal Research Scientist and Manager. AI giants are now joining forces with big telecoms to integrate AI early in the hardware design process, rather than adding it as an afterthought to existing devices.

This end-to-end collaboration may change the way people use their phones, says El Maghraoui. For example, users of Deutsche Telekom's AI Phone can simply press a button on the side of their phone and speak to activate the Perplexity assistant, who will make a dinner reservation, book a taxi or translate speech from one language to another in real time, all without opening the lock screen or any apps.

“They are redefining how users are interacting with their smartphones. The implications are huge,” says El Maghraoui.

This is not the first AI phone—or even the first AI phone boasting agents—and it won’t be the last. Earlier this year, Samsung announced it was bringing agentic AI to its Samsung Galaxy series of phones via Google’s Gemini assistant, which can now navigate between apps. A user activates Gemini by clicking a button on the side of the phone and can ask Gemini to find them flights cross-country and send them to a friend. Gemini would then search and compile flight options and pre-draft an email to that specific contact.

Google’s own Pixel phones offer Gemini as well, with agentic capacities since December 2024. A real-time AI assistant called Gemini Live lets people do things like take a photo of a plant and have a conversation about whether it will survive in their yard and when they should plant it.

Apple has been trying to catch up with its rivals and announced with much fanfare that it was unrolling Apple Intelligence features to its phones. Glitches, such as a news summarization tool that reversed the meaning of headlines, and delays have plagued Apple at several points. On March 7, an Apple representative confirmed in a statement: “We’ve also been working on a more personalized Siri, giving it more awareness of your personal context, as well as the ability to take action for you within and across your apps. It’s going to take longer than we thought to deliver on these features.”

Ultimately, the quality of the AI model will determine which AI phone wins, says Shobhit Varshney. He points to the fact that Google benefits from other businesses owned by its parent company (Alphabet), such as YouTube.

“They have billions of minutes of audio from YouTube to train their models on, much more than smaller AI players,” says Varshney.

Some, like researchers from Beijing University of Posts and Telecommunications, are now designing small language models specifically for smartphones. The team found that small language models run “more quickly and efficiently” when they are designed for the hardware they will run on rather than fine-tuning a larger model to use on a phone, says co-author Rongjie Yi to IBM Think.

Why is the AI phone race speeding up now? Advances in small language models (SLMs) are propelling this evolution in mobile technology, says El Maghraoui. “These compact, more efficient and more capable models can do real-time processing and create personalized, secure interactions on edge devices like smartphones,” she says.

El Maghraoui adds that as people use SLMs on smartphones, we may learn lessons that will actually be most useful outside the consumer arena. Small models are nice for phones, but critical for “more serious, mission-critical applications, such as embedded devices in factories where sensitive manufacturing data must be processed locally.”

From financial services to healthcare to manufacturing, there are many situations where companies have data that they want to keep on premises. Small models allow them to extract useful information from that data without having to send it to a public cloud.

For this reason, IBM  has prioritized smaller, fit-for-purpose models such as its Granite series because they enable enterprises to pursue frontier model performance at a fraction of the cost.

IBM’s Chairman and CEO Arvind Krishna elaborated in an article in Fortune: “In our work at IBM, we’ve seen that fit-for-purpose models have already led to up to 30-fold reductions in AI inference costs, making training more efficient and accessible.”

In many enterprise settings, there are “time and accuracy constraints that are more stringent,” says El Magrahoui. Any delays or errors can result in losing large amounts of money or compromising the security of data or more.

“You cannot afford to make a mistake in those environments because it has much more serious implications than just having maybe a mistake in scheduling a flight or something like that,” she says.

As users continue to experiment with SLM-powered AI-first phones, those in other arenas will continue to keep a close eye on what is learned from these edge devices.