On this page
You can use big open models like GPT-OSS 20B, Gemma 4 31B and Qwen 3.5 27B from your iPhone without paying per token, by running them on the Mac or PC you already own. Off Grid AI Mobile (OGAM), the free iPhone app, connects to Off Grid AI Desktop, Ollama or LM Studio on your Wi-Fi. The computer runs the model, your iPhone sends the question and shows the answer. There’s no API bill, because there’s no API company. It’s your hardware.
The steps use OGAM 0.0.124 from the App Store.
| Get OGAM for iPhone | Get Off Grid AI Desktop |
What your iPhone can run, and what it can’t
iOS gives one app only part of the phone’s RAM, and a running model needs room for the conversation on top of its file. OGAM measured a 12 GB iPhone 17 Pro Max loading a 5.6 GB Qwen 3.5 9B build at a 4K context. So OGAM sizes files like this:
| iPhone RAM | Runs well | Tight | Too big |
|---|---|---|---|
| 6 GB | under 3.6 GB | 3.6 to 4.2 GB | over 4.2 GB |
| 8 GB | under 4.8 GB | 4.8 to 5.6 GB | over 5.6 GB |
| 12 GB | under 7.2 GB | 7.2 to 8.4 GB | over 8.4 GB |
That covers models up to about 9B, such as Gemma 4 E4B, Qwen 3.5 4B and 9B, and Llama 3.2 3B. The models people pull most on Ollama at the bigger sizes are a different league:
| Model | Ollama download | Computer memory to aim for |
|---|---|---|
| GPT-OSS 20B | 14 GB | 24 GB |
| Qwen 3.5 27B | 17 to 20 GB | 32 GB |
| Gemma 4 31B | 19 to 20 GB | 32 GB |
| Qwen3 Coder 30B | 19 GB | 32 GB |
| Llama 3.1 70B | 43 GB | 64 GB |
Download sizes are from the Ollama library on 16 Nov 2026. The memory column adds room for the conversation and the system, so treat it as a rough guide. A 32 GB Mac or PC covers most of this list.
Why not just pay for an API?
You can. But per-token pricing means every long chat, every pasted document and every retry costs money, and your prompts go to someone else’s servers. Your own computer costs nothing per question. The only running cost is the electricity to keep it awake.
Step 1: Load the model on your computer
Use whichever you have. One is enough.
Off Grid AI Desktop. Open Models, download a model and select it. Ask it one question on the computer first. Its gateway listens on port 7878.
Ollama. Pull a model, for example ollama pull gemma4:31b. Then let other devices reach it. On a Mac:
OLLAMA_HOST=0.0.0.0 ollama serve
On Windows, add the system environment variable OLLAMA_HOST set to 0.0.0.0 and restart Ollama.
LM Studio. Download the model, open the Developer tab, load it, start the server and turn on Serve on Local Network.
Keep the computer awake and on the same Wi-Fi as your iPhone.
Step 2: Connect from your iPhone
- Open OGAM and go to Settings → Remote servers.
- Under Scan for, keep on Off Grid AI Desktop (7878), Ollama (11434) or LM Studio (1234).
- Tap Scan network. When iOS asks to let OGAM find devices on your local network, allow it. Without that, the scan can’t see your computer.
- Your computer shows up in the list. Turn on Auto-discover on Wi-Fi if you want OGAM to look again each time you open the screen.
If nothing shows up, tap Add manually. Give it a name and enter the computer’s address and port, such as http://192.168.1.50:7878. Tap Test connection, then Add server.
If Off Grid AI Desktop asks for a key, OGAM says so. Copy a key from Desktop’s Gateway → API keys and paste it into API key. It’s kept in the iPhone’s Keychain.
Step 3: Pick the model
Open a chat and tap the model name to open the model selector. Models on the iPhone are under Local Models. Models on your computer are listed under the server’s name. Tap Gemma 4 31B, or whatever you loaded, and ask.
Under Currently Loaded, a Remote badge tells you the answer comes from your computer. Switch to a local model in the same chat when you’re out.
Symptom, check, fix
| Symptom | Check | Fix |
|---|---|---|
| Scan finds nothing | Local Network allowed for OGAM? | iPhone Settings → Privacy & Security → Local Network, turn OGAM on |
| Still nothing | Same Wi-Fi? | Join the same network, or Add manually |
| Ollama missing | OLLAMA_HOST set? |
Set it to 0.0.0.0, restart Ollama |
| Server shows Not answering | Computer asleep? | Wake it, keep it awake |
| Desktop found, no models | API key | Paste a key from Gateway → API keys |
| Slow answers | Model vs computer memory | Use a smaller model from the table |
Using it away from home
On your home Wi-Fi this just works. From outside, join both devices to a private network like Tailscale; OGAM accepts Tailscale addresses. See using your home AI from anywhere with Tailscale. Don’t open these ports to the internet.
Need the full-size model?
Some models are too big for your computer as well. GPT-OSS 120B is a 65 GB download, and Qwen 3.5 122B is 81 GB. Off Grid AI Boost runs open models on our GPUs at one flat monthly price, from $4.99 for the first cohort (regular $9.99, plans up to $199.99). Still no per-token maths, and we don’t store your prompts or train on them. The full privacy policy comes with launch.
Boost is a pre-order. It starts in January; pre-order now to lock in half price, and you’ll get an API key. Usage limits for each plan aren’t published yet.
FAQ
Can an iPhone run GPT-OSS 20B? No. It’s a 14 GB download, and even a 12 GB iPhone tops out near 6 GB in OGAM. Run it on your computer and chat from the iPhone.
Is this free? OGAM, Off Grid AI Desktop, Ollama and LM Studio are free. Remote servers don’t need Pro.
Does my Mac need to be powerful? Memory matters more than speed. Aim for 24 GB for GPT-OSS 20B and 32 GB for Gemma 4 31B or Qwen 3.5 27B. With less, pick a smaller size of the same model.
Does anything go to the cloud? No. The iPhone talks to your computer over your own network. OGAM warns you if a server address is on the public internet.
What about when I have no Wi-Fi? Download a small model to the iPhone, such as Gemma 4 E4B, and switch to it. It runs with no connection at all.
Open Settings → Remote servers in OGAM and tap Scan network. Your computer’s models will be on your iPhone in a minute.