iFarmRun a local AI model to control an iPhone: which Mac you need
Memory decides it. A 16 GB Mac runs 7–8B vision models, 24–32 GB runs 12–24B models (the sweet spot), 48–64 GB runs 30B-class models, and 96 GB or more runs the largest open vision models. Run the model in Ollama or LM Studio, pick it in iFarm under AI brain → Custom / OpenAI-compatible, and screenshots never leave your Mac.
Why run the model on your own Mac
iFarm's auto mode lets an AI drive a real iPhone: it looks at a screenshot, decides, and taps, swipes or types through iFarm. The AI is the brain and iFarm is the hands, so the brain is your choice. A cloud model (Claude, OpenAI, Gemini) is the most capable. A local model has three different strengths:
- No subscription. No API key, no monthly plan, no per-use bill.
- No usage limits. No rate limits or quotas. Run it as often as you like.
- Nothing leaves the Mac. Screenshots and instructions stay on your Mac. No outside company sees them.
The cost is that the model runs on the Mac's own chip and memory, right next to the phones.
What the model has to be able to do
Not every model can drive a phone. It needs two abilities:
- See images. Every step starts with a screenshot of the phone. A text-only model can't tell where the search button is.
- Call tools (function calling). The model doesn't tap anything itself. It asks iFarm to tap, swipe or type by calling a tool. A model that can only write text back can't do that.
In Ollama's library, look for models with both the vision and tools tags. If a model page shows only one of them, it won't work as a brain.
Which Mac for which model
| Mac memory | Model size that fits | Example Macs | What to expect |
|---|---|---|---|
| 16 GB | 7–8B vision models | A base Mac mini M4 | Short, clear tasks: open an app, tap by name, read a value. A few seconds per step. Keep to about 4 phones. |
| 24–32 GB (recommended) | 12–24B models | A Mac mini M4 with more memory, or an M4 Pro | The sweet spot for a small farm, with room left for the phones. |
| 48–64 GB | 30B-class models | A Mac mini M4 Pro or a Mac Studio | The best local results for longer tasks. Pro chips also answer faster: more memory speed. |
| 96 GB and up | The largest open vision models | A Mac Studio with a Max or Ultra chip | A dedicated AI and farm machine. |
Remember the phones need memory too. Each phone costs the Mac about 140 MB (about 100 MB for its helper process and 40 MB for its live view). iFarm's Health page shows how much is free and "Room for about N more phones" for your Mac, so you can see what's left after the model loads.
Example models
New open models come out often, so treat these as starting points and check the library for newer ones. As of November 2026:
| Model | Sizes on Ollama | Vision | Tools | Notes |
|---|---|---|---|---|
| Qwen3-VL | 2b, 4b, 8b, 30b, 32b, 235b | Yes | Yes | qwen3-vl:8b for 16 GB; qwen3-vl:30b for 48–64 GB |
| Mistral Small 3.2 | 24b | Yes | Yes | Fits the 24–32 GB tier |
| Gemma 3 | 4b, 12b, 27b with vision | Yes | Check | Gemma 3 12B fits the 24–32 GB tier, but Ollama's official page lists only the vision tag. Use a build that's tagged for tools, or pick another model. |
Note: The tags on the model's own library page are what count. If iFarm's run stops at the first step, the model most likely can't call tools.
Set it up in iFarm (Ollama)
- Install Ollama from ollama.com.
- Download a model with vision and tools. In Terminal:
`` ollama pull qwen3-vl:8b ``
- In iFarm, open AI brain in the toolbar and choose Custom / OpenAI-compatible.
- Pick Ollama, press Check, and pick the model from the list.
- Open a phone's page, type a short task in its Claude box ("open Settings and read the iOS version") and press send (⌘↩). The box uses whichever brain you chose.
Set it up in iFarm (LM Studio)
- Install LM Studio and download a model that can see images and use tools.
- Start its local server in the Developer tab.
- In iFarm's AI brain, choose Custom / OpenAI-compatible, pick LM Studio, press Check and pick the model.
Anything else that speaks the OpenAI-style chat API works under Other.
Be realistic about what a local model can do
Local models are much weaker than Claude or GPT at driving a phone, and that's true even of the large ones on long, multi-step tasks.
| Works well | Struggles |
|---|---|
| Open an app | Long tasks with many screens |
| Tap something by its text | Tasks that need judgment on every screen |
| Read a value off the screen | Recovering after a wrong turn |
| Short, clearly worded goals | Vague goals ("do something useful") |
They're also slower: on a 16 GB Mac, expect a few seconds per step. Bigger models do better, which is why memory matters.
The trick: let the AI run once, then replay for free
For jobs you repeat every day, you don't need an AI at all. A recorded automation is faster and free. So use the local model to build the automation:
- On a phone's page, type the task in its Claude box.
- Turn on Save as automation (save what the AI does as a new automation).
- Let the model do it once while you watch.
- iFarm saves every step it took as a normal automation. From then on it replays without any AI: fast, free, on any phone, on schedules and in flows.
If the model took a detour (tapped the wrong thing and went back), the saved steps include it; delete those steps in the editor. Each saved tap gets the same text matching as a recording, so it also works on other iPhone sizes (how that works).
The rules still apply
A local brain doesn't change what iFarm allows. By default iFarm asks before posting, sending or following, and you approve each post. iFarm never buys anything, never types passwords, and doesn't hide automation from apps. Stop, ⌘. and the phone's emergency stop (hold volume up + volume down) work the same. Each app's own rules still apply whichever brain you use.
Frequently asked questions
Can a 16 GB Mac mini run a local model and phones at the same time?
Yes, for small models and a small farm. A 16 GB Mac fits 7–8B vision models such as qwen3-vl:8b; keep to about 4 phones and short, clear tasks.
Is a local model as good as Claude for phone automation?
No. Local models are fine for short tasks but often get lost on long or judgment-heavy ones. For daily routines, record an automation, or let the model build one once and replay it without AI.
Do screenshots leave my Mac with a local model?
No. With Ollama or LM Studio on the same Mac, screenshots and instructions stay on the Mac.
Why does my model start and then do nothing?
It most likely can't call tools, or can't see images. Pick a model with both the vision and tools tags in Ollama's library.
Does iFarm cost anything with a local model?
No. iFarm is free with every feature, and a local model has no per-use bill.