You donβt need a new phone, a powerful PC, or the cloud. A private AI β one that answers your questions offline, with no account and no one watching β can run on the old device gathering dust in your drawer. Hereβs the easiest free way for whatever you own. ![]()
Start here β pick your device
| What you have | What to do |
|---|---|
| llamafile β one file, double-click, chat. No install. | |
| PocketPal β a free offline chat app, no root needed. | |
| WebLLM β a model runs right in the tab. | |
| llamafile on Pi β the same one-file trick. |
The honest bit first: RAM is the real limit β the model plus a little breathing room must fit. Old cheap phones only run small models; a very old iPhone/iPad on old iOS genuinely cannot without a jailbreak. Answers are slow (a few words per second), not desktop-fast. Keep that in mind and the rest is all upside.
What made this possible
smolmaxxing is a free, MIT project that proved a 16B AI brain can run on a 1 GB TV box and a 2013 iPad β by patching llama.cpp to keep the busy parts of a model in RAM and read the rest from storage as needed. Every number was measured on the real device.
| Device | Everyday model | Biggest model |
|---|---|---|
| Qwen3.5-2B @ 7.7 tok/s | Ling-mini 16B @ 5.1 tok/s | |
| Bonsai-1.7B @ 24β26 tok/s | Ling-mini 16B @ 2.6β4.5 tok/s | |
| Qwen3.6-35B @ ~5 tok/s | Qwen3-Next 80B @ 2.8 tok/s |
Full logs: TV box Β· iPad Β· phone Β· 160 experiments
The free toolbox, by device
| Device | Tools (all free) |
|---|---|
| llamafile Β· llama.cpp builds Β· LM Studio | |
| PocketPal Β· ChatterUI Β· Maid Β· Termux | |
| PocketPal Β· Google AI Edge Gallery (recent devices) | |
| WebLLM Β· transformers.js | |
| Pi llamafile guide |
Ungamed rare gems
NIGHTRUN boots a chat AI straight from a USB stick with no operating system at all.
BigMoeOnEdge streams a ~284B model on a 12 GB phone by reading only the parts it needs β needs fast storage and ~91 GB free.
llamita.cpp runs an 8B 1-bit model on a $35 Pi; project-willamette brings 1-bit AI to old Pentium-class CPUs.
Not sure what fits? FitLLM and LocalLM Advisor take your specs and name the biggest model you can run.
GGUF-my-repo turns any model into a small free file right in your browser; awesome-mobile-llm is a catalog of tiny models.
The same idea, pushed to the absurd
| Project | The extreme |
|---|---|
| PFor | 180M params on a $6 ESP32 chip. |
| ps2-llm | A model on the PlayStation 2. |
| iiama2.c | On a Nintendo Wii, 61.9 tok/s. |
| esp32-ai | 28.9M params on an ESP32-S3. |
| LLM-on-anything | The whole hidden menagerie. |
Make it actually useful
Talk to it with whisper.cpp; talk back with Piper or eSpeak-NG.
Serve it to every device at home via llama.cpp web UI, or chat with your own files using sqlite-vec.
Keep it on with Node-RED or offline Wikipedia via Kiwix; a 3 MB HermitUI page and a single-file llamafile reach it from any old machine.
Dig deeper β the original build notes
How it works Β· what didnβt work Β· all model files Β· TV box setup.
The 1-bit frontier: vlut.cpp (paper) Β· BitNet 1.58-bit (GGUF) Β· Ternary Bonsai.
Verified prebuilt Android binaries; the old-iPad route used checkra.in.
Fine print: Speeds are author-reported and headless; tiny models can ramble. Model files carry their own licences, not MIT. The jailbreak is only needed for the very old iPad experiment β every βstart hereβ path above avoids it. Got a small device? Submit your numbers.
Free & open, built by MohammedRashad / Hyperion AI (Nov 2026). ![]()

!