1
Pick a model that fits
Seven open models from 0.8 to 9 billion parameters, each with a “Fits your device” badge read off your phone's RAM. Download once; pause and resume if the Wi-Fi drops.
Personal LLM
Free on iOS & Android
Private AI that
never phones home.
Personal LLM runs open models — Qwen 3.5, Gemma 4, GLM, Ministral — on your phone's own chip. Download one once, then chat, ask about photos and watch it reason with no connection at all. There is no server. Your conversations have nowhere to go.
no account
works offline
understands photos
open-weight models
no subscription
The one idea
Everything interesting happens on the phone.
Cloud chatbots send every word to a data centre and hope you trust them. Personal LLM doesn't ask you to: the model is a file on your phone, the answer is computed by your phone, and the only things that ever touch the network are the two below.
Your phone
- Your chats stay
- Photos you ask about stay
- The models themselves stay
- Settings & system prompts stay
- Usage stats stay
- Backups you export yours
Computed here, stored here, deleted when you say. Uninstall and it's all gone — nothing was ever anywhere else.
Model file · once, from Hugging Face
Ads in the free app · only while online
Chats, photos, prompts · never
The internet
- Hugging Face model downloads
- Google AdMob ads
- Our servers there aren't any
After the download, switch on airplane mode if you like. The app won't notice. Ads can't load without a connection — everything else carries on.
How it works
Download once. Talk forever.
Three screens you'll actually see, in the order you'll see them.
2
Chat anywhere
On a plane, in a basement, in a tunnel. Answers stream in with rendered Markdown and a live tokens-per-second readout, and nothing is waiting on a signal.
3
Ask about a photo
Every catalog model can take an image. Add the image-support file once (195 MB–1 GB) and point it at a menu, a form, a whiteboard, a sign in a language you don't read.
The catalog
Open weights, phone sizes.
Quantized GGUF builds that run through llama.cpp — Metal on iPhone, OpenCL on Snapdragon Adreno, CPU everywhere else. Tell us your phone's RAM and we'll badge them exactly the way the app does.
Qwen 3.5 0.8B
For very low-end phones. Thinking mode on; quality is limited, speed is not.
Gemma 4 E2B
Google DeepMind's phone-first model, quantization-aware trained. 140+ languages.
Ministral 3 3B
Mistral's compact model with a 256k context window — quick on documents and screenshots.
Qwen 3.5 4B
The sweet spot: speed, quality and thinking mode on an everyday phone. Our recommended first download.
Gemma 4 E4B
The higher-quality Gemma, with hybrid thinking for harder questions.
Qwen 3.5 9B
The best answers in the catalog. Thinking mode for math, code and anything you'd like to check.
GLM 4.6V Flash
Zhipu AI's flagship-class vision model: visual Q&A, documents, UI screenshots.
Bring your own
Paste a link to any GGUF and it joins the catalog, with an optional vision projector for image input.
Same rule as the app: “Fits your device” means at least 2 GB of headroom over the model's minimum; “Should run” means it meets it. Image support is a separate 195 MB–1 GB download per model.
Thinking mode
Watch it reason, then fold it away.
On reasoning models, switch on thinking mode and the step-by-step working shows up in a collapsible Reasoning panel above the answer — handy for math, code and anything worth checking. It's off by default, because most questions don't need it and it's faster that way.
- One-tap presets — Creative, Balanced, Precise, Thinking
- Temperature, top-p, top-k, context size, max tokens — all yours to tune
- A system prompt per chat — the recipe assistant doesn't have to sound like the code reviewer
- Edit, regenerate, stop — a stopped reply keeps what it had
Creative t 1.2
Balanced t 0.7
Precise t 0.3
Thinking reasoning on
Built for phones
A real chat app, not a tech demo.
Running a language model on a phone is the clever part. The rest of Personal LLM is the unglamorous part — the stuff that makes it the app you actually open.
GPU acceleration
Metal on iPhone, OpenCL on Snapdragon Adreno 700+, via llama.cpp. Falls back to CPU quietly when there isn't one.
Honest numbers
Tokens per second on every reply, context usage in the header, and a Fits-your-device badge before you download.
Markdown that renders
Code blocks, tables and lists come out formatted. Copy any reply in one tap.
Switch models mid-chat
Start on the fast one, hand a hard question to the 9B from the chat header. Same conversation.
Read aloud
On-device text-to-speech for any answer — no voice service, no upload.
Backup & restore
Every chat and setting as one JSON file, yours to keep. Restore from the file or the clipboard.
Downloads that survive
Pause, resume, cancel. Interrupted downloads pick up where they stopped, and storage is checked before they start.
Light and dark, search and pin
Follows your system theme or pick one. Search every chat, pin the ones you come back to.
Questions
The short answers.
Is it free?
Yes. There's no subscription and no paywall on any feature — the models are free open weights and the app costs nothing. It's supported by ads: a small banner in chat and an occasional full-screen ad when you open a new chat or while a model downloads, never in the middle of an answer. Offline, none load.
Does anything leave my phone?
Your chats, the photos you attach and the models themselves never do — inference runs on the phone's own chip and we don't operate a server. Two things use the network: downloading a model file (once, straight from Hugging Face) and ads. Uninstall the app and everything is gone, because nothing was ever anywhere else.
Which phone do I need?
iOS 15.1 or later, or Android 7 or later. Small models run in 3 GB of RAM; the 9B models want 6–8 GB. GPU acceleration uses Metal on iPhone and OpenCL on Snapdragon Adreno 700+ phones, with a CPU fallback everywhere else. Every model card shows a “Fits your device” badge read off your actual RAM before you download.
How much storage do the models take?
From 0.81 GB (Qwen 3.5 0.8B) to 6.17 GB (GLM 4.6V Flash), plus 195 MB–1 GB if you add image support to a model. Downloads can be paused and resumed, and the app checks free space before it starts.
How fast is it?
That depends on the phone and the model — a 4B model on a recent phone streams at reading pace; a 9B is slower and smarter. Every reply shows a live tokens-per-second readout so you can compare models on your own hardware, and you can switch models mid-conversation from the chat header.
Can I bring my own model?
Yes. Add any GGUF by URL — Hugging Face links work — with an optional vision projector for image input. Temperature, top-p, top-k, context size, max tokens, thinking mode and per-chat system prompts are all adjustable, or use the Creative, Balanced, Precise and Thinking presets.
What about ads and tracking?
Ads come from Google AdMob. On iOS the app asks for App Tracking Transparency permission; decline it and ads are simply non-personalized. AdMob may collect device information to serve ads — that's spelled out in the privacy policy. The app itself has no account, no cloud sync and no analytics on your conversations.
The smartest thing on your phone can also be the most private.
Free. No account. Works on a plane.