Phone AI began in the cloud. Early assistants sent your voice to a company’s servers because phones couldn’t do the work themselves. Around 2017, phone chips gained dedicated AI hardware, and narrow tasks such as face unlock, photo processing and keyboard suggestions moved onto the device. Since late 2023, generative models small enough for phones have arrived, and today’s phone AI is a hybrid: a small model on the device for quick, personal tasks and a large one in a data center for the rest. Privacy got better for everything that moved onto the phone, but not every feature labeled “AI” is local.
A short timeline of phone AI #
| When | What happened | Where the AI ran |
|---|---|---|
| 2011 | Siri launches on the iPhone 4S | Cloud |
| 2017 | Apple’s A11 Bionic chip adds a Neural Engine, powering Face ID; Huawei’s Kirin 970 ships with a dedicated neural processor | On the device, for specific tasks |
| Late 2010s | Speech recognition, photo enhancement and some translation start working offline on many phones | On the device, for specific tasks |
| 2023 | Open-source tools like llama.cpp make it practical to run quantized open language models on laptops and phones | On the device, general chat |
| December 2023 | Gemini Nano, Google’s on-device model, ships on the Pixel 8 Pro | On the device, generative |
| January 2024 | Samsung launches Galaxy AI with the Galaxy S24, including Gemini Nano | Hybrid |
| 2024 | Apple Intelligence arrives for the iPhone 15 Pro and later, with Private Cloud Compute for bigger requests | Hybrid |
| August 2025 | Pixel 10 introduces Magic Cue, proactive suggestions using on-device machine learning | Mostly on the device |
| August 2026 | Pixel 11 brings “Gemini Intelligence,” using cloud Gemini and on-device Gemma models | Hybrid |
Era 1: everything in the cloud #
The first mainstream assistants were thin clients. Your phone recorded what you said, sent it to the company’s servers, and read back the answer. Phones didn’t have the processing power or memory to understand speech or language on their own.
The privacy cost was simple: everything you asked went to the company, often along with your location and contacts to make the answer useful.
Era 2: narrow AI on the chip #
Dedicated neural hardware changed what phones could do locally. Face unlock, portrait-mode photos, scene detection, keyboard predictions and offline dictation all moved onto the device, because running them locally was faster and didn’t need a connection.
This era brought the first big privacy wins. Biometric data like your face or fingerprint could stay on the phone. Photo processing didn’t require uploading your pictures. Keyboards could learn your typing habits without sending every keystroke off the phone. We cover that last one in is your keyboard private.
Era 3: generative AI on the device, and the hybrid model #
Language models used to be far too big for phones. Three things changed:
- More memory. Flagship phones moved to 8 GB of RAM and more, enough to hold a model with a few billion parameters.
- Better chips. Phone GPUs and neural processors became fast enough to generate text at reading speed.
- Smaller models. Techniques like quantization, which stores a model’s numbers in fewer bits, and distillation, which trains a small model to imitate a large one, shrank capable models to a few gigabytes. GGUF quantization explained shows what that looks like in practice.
The result is the hybrid pattern Apple, Google and Samsung all use now: on-device models for summaries, writing help and suggestions, and cloud models for anything bigger. Apple’s version routes large requests to Private Cloud Compute; Samsung lets you choose on-device-only processing; Google’s newest system mixes on-device Gemma with cloud Gemini.
Alongside the built-in features, open models from Google, Alibaba, Mistral and others made full chat apps possible on phones. Personal LLM is one example: it runs Qwen 3.5, Gemma 4, Ministral 3 or GLM 4.6V Flash entirely on the phone, on iPhones from iOS 15.1 and Android phones from Android 7, so on-device AI isn’t limited to the newest flagships. After a one-time model download it works offline, and chats never leave the phone.
What the shift means for your privacy #
Moving AI onto the phone is the biggest privacy improvement in mobile AI so far, with three caveats.
- Hybrid features still use the cloud. Many “on-device” systems send larger or harder requests to servers, and the phone decides which path a request takes.
- Local processing doesn’t mean no data collection. An app can process your request on the device and still send analytics, or draw on data already stored in a cloud account, like email.
- Labels are loose. “AI-powered” and “on-device” in marketing don’t tell you where a particular feature runs.
Two habits help. Check your phone’s AI and data settings (our guide to phone AI privacy settings lists them for iPhone, Pixel and Galaxy). And use the airplane-mode test: if a feature still works with every connection off, it’s running on the device. For a fuller explanation of what stays on the phone, see is on-device AI private.
What still needs the cloud? #
- The biggest, most capable models. Hard reasoning, long documents and complex code still run best on servers.
- Anything current. Web search, news, weather and live information need a connection.
- Some creative features, like high-quality image generation, on most phones.
Where things go next, including proactive and agentic assistants, is covered in the future of AI assistants on your phone. If you’re deciding whether a new phone is worth it for AI, see are AI phones worth upgrading for.
Frequently asked questions #
When did phones start running AI on the device? #
Specific AI tasks moved onto phones around 2017, when chips such as Apple’s A11 Bionic and Huawei’s Kirin 970 added dedicated neural hardware for things like face unlock and photo processing. Generative language models on phones arrived in late 2023, with Gemini Nano on the Pixel 8 Pro.
Is on-device AI more private than cloud AI? #
Generally yes, because your data doesn’t have to leave the phone to be processed. But hybrid systems still send some requests to servers, and an app can process locally while collecting other data, so check each feature’s settings.
Why do phones still use the cloud for AI? #
The most capable models are far too large for a phone’s memory, and anything involving current information needs the internet. Phones handle quick, personal tasks locally and hand the rest to servers.
Do I need a new phone for on-device AI? #
Not for everything. Built-in features like Apple Intelligence and Gemini Nano need recent phones, but apps that run open models can work on phones several years old, with smaller models on phones that have less memory.
What’s the difference between edge AI and on-device AI? #
On-device AI runs on the phone itself. Edge AI is a broader term for processing near where data is created, which can mean your phone or a nearby device or server rather than a distant data center.