5 native AI apps that lastly make working an LLM on my cellphone value it

My first strive at native AI on a cellphone was a bare-bones chat app again in 2024. I loaded a tiny 1B mannequin, requested it a handful of questions, and received gradual, shaky solutions. It lasted a couple of week earlier than I deleted it. ChatGPT was at all times one faucet away, and it was smarter.

A drive via Central Pennsylvania introduced me again. Long stretches exterior State College drop my sign to nothing, and I had a pile of voice notes I wished cleaned up earlier than attending an internet assembly. No cloud mannequin may assist.

The small fashions are a lot better now. Qwen 3.5, Gemma 4, and Liquid AI’s LFM 2.5 all match on a cellphone, and Apple’s personal on-device mannequin is open to third-party apps via the Foundation Models framework. The larger shift for me is within the apps. A couple of builders have wired these fashions into Siri, Shortcuts, and the Action Button, so I can attain them with out opening a chat window first.

I examined all the pieces under on an iPhone 17 Pro Max working iOS 27. These 5 made the reduce for my dwelling display screen. In the tip, like other local LLMs, I discovered them every helpful.

Locally

LM Studio’s iPhone app can borrow your Mac’s horsepower

Locally started as an indie app from developer Adrien Grondin. LM Studio, the desktop app many people use to run models on their computer systems, bought it in April 2026, and it now serves as LM Studio’s official iPhone app.

Locally runs fashions proper on the cellphone utilizing Apple’s MLX framework. No account is required, and the App Store itemizing says the developer doesn’t gather any information. Available fashions embrace Qwen 3.5, LFM 2.5, Gemma, Llama, and Apple’s personal Foundation mannequin. Qwen 3.5 may also deal with photos. You can open Voice mode from a Home Screen fast motion, a Shortcut, or a button in Control Center. I largely used Qwen 3.5 on the 2B measurement, and replies began virtually instantly. When I gave it a photograph of a receipt, it discovered the full with none bother.

LM Link is what stored me utilizing it. Released in June, it pairs the app with LM Studio working on a house laptop and makes use of end-to-end encryption. My MacE book Air does the heavy lifting, so my cellphone will get an even bigger mannequin with out the warmth or battery drain of working it regionally. Chats keep saved on the system. One caveat: on this mode, the cellphone is a distant, so the pc does the processing.

I loaded Qwen 3.5 9B on my 16GB MacE book Air and gave it the identical immediate I’d simply run on the cellphone’s 2B mannequin: a top level view evaluating NFL cities with the groups that decision them dwelling. The LM Link model was slightly slower to reply, although I needed to look ahead to it. Both outlines had been thorough, they usually took the subject in numerous instructions. I preserve flipping between the 2 in the identical app and evaluating solutions.

Private LLM

A $4.99 app that plugs native AI into Shortcuts and Siri

I attain for Private LLM after I need AI inside an automation. It prices $4.99 as soon as, with no subscription, and Family Sharing covers up to six people.

Private LLM works with Siri and the Shortcuts app, so you need to use a mannequin as one step in an automation as a substitute of opening the app and typing right into a chat. It runs fashions via an MLC-LLM engine with OmniQuant quantization. According to the developer, that retains solutions sharper than the extra primary compression different apps use. The mannequin record consists of Qwen 3, Phi 4, and Gemma, and newer Pro iPhones may also run Llama 3.1 8B. It solely works with textual content, although. You can’t give it a photograph.

I take advantage of dictation quite a bit on my iPhone, and I normally want to wash up the textual content earlier than I can use it. My Shortcut sends no matter I dictate to Qwen 3 4B with directions to repair the punctuation and add paragraph breaks, however depart my wording alone. I put it on the Action Button. In my testing, what got here again was clear sufficient to stick straight into an e-mail.

I half anticipated lengthy pastes to journey it up. I dropped in a 750-word article and requested for a abstract, and it dealt with it with out stalling.

PocketPal AI

Any Hugging Face mannequin, downloaded straight to your cellphone

The free PocketPal AI is open source, and it is constructed on llama.cpp. Its enchantment is vary. You can search Hugging Face from contained in the app, decide any GGUF mannequin, and select a quantization stage that matches your cellphone’s reminiscence. If a mannequin is gated, paste in a Hugging Face entry token, and also you’re via.

That makes it the app I open when a brand new small mannequin drops, and I wish to strive it the identical afternoon. PocketPal additionally has “Pals”, saved assistants with their very own system prompts, plus a group hub for locating others people-made. The itemizing says it collects no information, and the app helps iOS 15.1 and later, so it is the friendliest possibility for an older iPhone.

I pulled Gemma 4 E2B straight from Hugging Face, a roughly 4GB obtain. Once it loaded, I requested it to elucidate the distinction between a 2-billion and a 9-billion-parameter mannequin, and it gave a radical reply on display screen inside just a few seconds.

Google AI Edge Gallery

Google’s Gemma 4 playground comes with instruments in-built

Google’s entry is much less of a each day assistant and extra of a check bench for its Gemma fashions. It’s free, requires iOS 17 or later, and now helps Gemma 4, together with a Thinking Mode that reveals the mannequin working via an issue.

Agent Skills is the half value attempting. A talent offers the mannequin a instrument, resembling Wikipedia lookups for fact-checking, an interactive map, or a visible abstract card. You can load extra abilities from a URL. Google labels the app as in energetic growth, and it reveals within the tough edges. Don’t confuse it with the Gemini app, which is a separate product.

I attempted the Wikipedia talent (it wants a connection) on just a few fact-check questions on “Friends.” Gemma 4 pulled the precise web page every time, and the summaries had been correct. With Thinking Mode, a mannequin setting, switched on, I received roughly the identical info in a distinct format, and the velocity barely modified. This is a toggleable characteristic contained in the AI Chat module that exposes the mannequin’s inside reasoning course of on to the person.

Enclave AI

Chat together with your PDFs, even in Airplane Mode

Enclave is free, with in-app purchases for bigger fashions, and runs on llama.cpp. It handles two jobs I care about. Local voice chat works with no web connection, and you’ll work with paperwork straight within the app. There’s additionally an “Ask Enclave” Shortcuts motion and Siri assist.

Enclave may also connect with cloud fashions resembling Claude and Gemini. That’s helpful, however verify which mannequin is chosen earlier than you paste in something delicate.

I loaded Google’s Gemma 4 E2B (an 8GB obtain) in Enclave and fed it a PDF. Answers got here a bit slower than within the different apps, however they had been correct and pulled from the precise components of the doc. Voice chat works too. I turned on Airplane Mode and talked to it with no connection, however replies lagged, and the voice sounded extremely robotic.

How to choose your first mannequin

Parameters, quantization, and RAM in plain English

Model lists in these apps are complicated at first. You get names like Gemma 4 E2B and Qwen 3.5 9B, and file names ending in issues like Q4_K_M. I ignored all of it at first and tapped no matter was on the high of the record. That’s high-quality till you decide one thing too massive on your cellphone.

Look on the quantity with a B after it first. It’s the parameter rely in billions, and you’ll consider it loosely as how a lot the mannequin realized throughout coaching. More parameters typically get you higher writing and broader information, however they want extra reminiscence and take longer to reply. I attempted Qwen 3.5 at two sizes. The 2B model on my cellphone responded virtually straight away. The 9B model working on my Mac took slightly longer and went in a distinct path. Both had been helpful. If you’re new to this on an iPhone, stick to one thing between 2B and 4B.

Then there’s quantization, which is principally compression for AI fashions. Browse Hugging Face in PocketPal, and also you’ll see file names ending in Q4_K_M, Q5_K_M, or Q8_0. The quantity after the Q is roughly what number of bits each bit of the mannequin retains. Q8 is closest to the unique and takes probably the most area. This autumn is about half the scale and loses lower than you’d anticipate. For a cellphone, Q4_K_M is normally the one to seize. It’s additionally why the identical Gemma 4 E2B mannequin was a 4GB obtain in PocketPal and an 8GB obtain in Enclave.

Last, verify your reminiscence. Apple does not put RAM on its spec sheets, so lookup your mannequin earlier than you obtain something massive. iOS and your different apps want room too, so decide a file just a few gigabytes smaller than your cellphone’s complete reminiscence. The 8GB Gemma ran on my 12GB cellphone, nevertheless it was the slowest of all the pieces I examined.

Keep an eye fixed on storage, as properly. Every iOS app retains its information separate, so if you happen to strive three apps, you obtain the identical mannequin 3 times. I didn’t take into consideration that till I checked my storage and located Gemma 4 sitting on my cellphone in 4 completely different locations, consuming up g

Local fashions additionally know lower than ChatGPT or Claude, they usually haven’t any reside net entry until an app provides a instrument for it. I nonetheless use my paid subscriptions for analysis. Local now covers the short stuff for me, like cleansing up notes and something I’d quite not add.