AI Inside Mobile Phone

Ahmad Hakimi Adnan

6 min read

·

Sep 25, 2026

--

Listen

Share

Phones typically fight on specs such camera, processor and other, but they also fight on whose AI is smarter. So forget the specs. Let’s talk about the actual AI.

Basics Information

Mobile AI works two ways:

1. On-device model: a small AI that lives inside the phone. Fast, works offline, private.

Example:

• Photos app (iPhone) / Google Photos: Type “cat” or “bald” in the search bar and it finds all your cat or bald photos, even though you never labeled them.
• Camera app (iPhone Live Text): Point the camera at a sign or menu, tap the text icon, and copy the words straight from the picture.
• Notes app: Tap the mic icon on the keyboard and talk, and it types what you say.
• Keyboard: Type “See you” in WhatsApp and it suggests “tomorrow.”
• Face ID / Face Unlock: The phone recognizes your face to unlock.

2. Cloud model: a big AI on a server. Smarter, but needs internet and your data leaves the phone.

Examples:

• ChatGPT app: Type “Write a leave email.” and it writes the whole email.
• Gemini app: Ask “Plan a 3-day trip to Langkawi under RM 20” and it gives you a full plan.
• Image Playground (iPhone): Type “cat wearing sunglasses on a beach” and it creates the picture.

Every brand now uses both. The small one handles quick stuff, the big one handles hard stuff.

1. Apple: Siri AI

Timeline

• 2011: Siri first launched on the iPhone 4S.
• June 10, 2024: Apple introduced Apple Intelligence and a redesigned Siri that would understand our personal context, act across apps and respond to what’s on screen. Source: OpenClassActions
• March 7, 2025: Apple delayed the smarter Siri indefinitely, saying features like using personal info and controlling apps would come “in the coming year.” Source: MacTech
• March 20, 2025: Apple got sued for false advertising over the delayed features. Source: Axios
• May 2026: Apple agreed to pay $250 million to settle. Source: The Apple Post
• June 8, 2026: Apple revealed AFM 3, the new AI models behind Siri. Source: Apple ML Research
• Sept 14, 2026: Siri AI rolled out in beta with iOS 27. Source: ToKnow.ai

Before: what old Siri could do

• Set alarms and timers
• Call or text someone
• Check the weather
• Play music

The 2024–2025 period
Apple showed a demo where someone asked Siri about their mom’s flight and lunch plans using info from Mail and Messages, and advertised it in a TV ad starring Bella Ramsey. Source: MacRumors

But it didn’t ship. Apple delayed it in March 2025 and pulled the ads, after they’d already been running for months. Source: Fortune

Apple even shook up the leadership team in charge of Siri. Source: Yahoo Finance

Result: US buyers of the iPhone 16 and iPhone 15 Pro can now claim $25 to $95 per phone, with claims closing December 21, 2026. Source: Clarkson Law Firm

Now: the AI inside the phone (AFM 3)

Siri AI runs on two local models on the phone:

• AFM 3 Core: a 3-billion-parameter model for everyday tasks
• AFM 3 Core Advanced: a 20-billion-parameter model that only switches on 1 to 4 billion parameters at a time Source: SoftwareSeni

Core Advanced is multimodal, meaning it can see (images, screen), not just read text. It powers advanced dictation and expressive voices. Source: MacStories

The improvement
Apple calls AFM 3 Core the next generation of its 3-billion-parameter model. Source: Apple ML Research

So before, the biggest AI inside the iPhone was 3 billion parameters. Now it’s 20 billion. That’s almost 7x bigger.

What Siri can do now (on the phone)

• Reads your screen
  Siri understands what’s on your screen, so you can say something like “add this invoice to my expenses” and it knows what “this” means. Source: SoftwareSeni
• Knows your stuff
  It pulls context from your messages, email, photos and calendar. Source: SoftwareSeni
• Does multi-step tasks
  Apple’s own example: Siri finds an activity mentioned in a message, grabs a related recipe from your email, and adds the ingredients to Reminders. Source: AiCybr
• Works with apps
  It can send a WhatsApp message or play an Audible book, with deeper support coming for apps like Microsoft Outlook. Source: AiCybr
• Sounds human
  Siri’s new expressive voices are generated in real time, fully on the phone. Source: Apple ML Research
• Works offline
  Screen awareness, cross-app context and everyday actions don’t need internet. Source: SoftwareSeni

Apple trained the models on its own data, then refined them with outputs from Google’s Gemini. Apple says Siri AI is not “powered by Gemini” and there’s no Gemini in iOS. Source: MacStories

The problem

• Beta only, and not in the EU at launch Source: Apple Newsroom
• Apple hasn’t benchmarked its top model against GPT-6 Astra or Claude, and nobody positions it as a frontier model Source: andrew.ooo

Siri AI vs Other Phone AIs

Apple iPhone: Siri AI

• AFM 3 Core (3 billion parameters) and AFM 3 Core Advanced (20 billion parameters, uses 1 to 4 billion at a time) Source: SoftwareSeni
• Built by Apple, refined with Gemini outputs Source: MacStories

What it can do:

• Reads your screen: “add this invoice to my expenses” Source: SoftwareSeni
• Finds an activity in a message, grabs a recipe from email, adds ingredients to Reminders Source: AiCybr
• Sends WhatsApp messages, plays Audible books Source: AiCybr
• Human-sounding voices, generated fully on the phone Source: Apple ML Research

Google Pixel 11: Gemini Intelligence

https://www.android.com/ai/

Model used:

• Gemini Nano 4, Google’s newest on-device model, and Pixel 11 is among the first phones with it Source: 9to5Google
• Built fully by Google
• Before: Pixel 10 used Nano v3 Source: tech-ish

What it can do:

• Shows helpful info inside apps before you ask Source: TechRepublic
• Pulls your flight from Gmail and tells you if it’s delayed Source: TechRepublic
• Dubs foreign videos, podcasts and voice messages live, on the phone Source: TechRepublic
• Turns sign language into text through the camera Source: TechRepublic

Compared to Siri:

• Model: Siri uses Apple’s own AFM 3. Pixel uses Google’s own Gemini Nano 4. Apple shared its model size (20B), Google focuses on speed: up to 3.5x faster on-device AI. Source: Thurrott
• Style: Siri acts when you ask. Pixel acts before you ask.
• Unique to Pixel: Live video dubbing and sign language to text. Siri doesn’t have these.
• Unique to Siri: Screen-based commands like “add this to my expenses.”

Samsung Galaxy S26: Galaxy AI

Model used:

• Google’s Gemini Nano on the phone, since the S26 meets the Nano v3 or newer requirement Source: tech-ish
• Bixby rebuilt with a new AI model powered by Perplexity Source: Android Faithful
• Built by Samsung + Google + Perplexity. Samsung doesn’t have its own main AI brain.

What it can do:

• Bixby changes settings from normal talk: “battery is at 30%, make it last longer” Source: FN News
• Gemini books a taxi, you just check and tap confirm Source: Samsung Newsroom
• AI answers unknown calls and asks who’s calling and why Source: The Deep View

Compared to Siri:

• Model: Siri runs on one AI made by Apple. Samsung mixes three: Bixby, Gemini and Perplexity.
• Style: Siri is one assistant for everything. Samsung makes you pick which AI to use, and reviewers say the three don’t talk to each other. Source: Android Faithful
• Unique to Samsung: AI call screening and a choice of assistants.
• Unique to Siri: Apple’s own on-device model and a single assistant that works across your apps.

Summary

• Apple: Own AI (AFM 3), biggest shared on-device model (20B), knows personal stuff
• Google: Own AI (Gemini Nano 4), fastest and most proactive
• Samsung: Borrowed AI (Gemini + Perplexity), most choices, but messy