Cloaked ships 11 open-weight models from 5 labs, ranging from 317MB to 5.6GB. Qwen 3.5 0.8B is the recommended default for most people — small, fast, vision-capable, and light enough that heavy turns stay clear of the memory ceiling. If your device has the memory for it, Qwen 3.5 4B is the strongest all-around model available. For coding, Phi-4 Mini is the specialist.
All of them run on-device via Apple’s MLX framework, and none of them transmit anything. For background on how that works, see The Complete Guide to On-Device AI, How LLMs Actually Run on Your iPhone, and What Is Apple MLX?.
Read the RAM column, not the size column
The most common mistake is choosing a model by download size. Required memory is what the app actually enforces, and it is much larger than the file.
Weights have to be resident, but so does the key-value cache — which grows with conversation length — plus the vision encoder and image tensors on multimodal models, plus working memory. The stated minimum covers the peak of a heavy turn, not the idle case.
That’s why a 2.9GB model can require 10GB of RAM. Which iPhones Can Run Local AI maps the requirements onto specific devices.
Token speeds below are the catalogue’s estimates for an iPhone 15 Pro. Real speeds vary with thermal state, conversation length, and background load — expect 25–40% lower on a warm device. For context, comfortable reading is about 5–6 tokens per second, so nearly everything here outruns reading. Tokens Per Second Explained covers why.
The full catalogue
| Model | Provider | Size | Min RAM | Context | Speed | Vision | Thinking |
|---|---|---|---|---|---|---|---|
| Qwen 3 0.6B | Alibaba | 317MB | 4GB | 32K | ~45 | — | Toggle |
| Qwen 3.5 0.8B ★ | Alibaba | 622MB | 5GB | 32K | ~40 | ✓ | Toggle |
| Qwen 3 1.7B | Alibaba | 914MB | 4GB | 32K | ~30 | — | Toggle |
| Llama 3.2 1B | Meta | 800MB | 4GB | 8K | ~28 | — | — |
| Qwen 3.5 2B | Alibaba | 1.6GB | 6GB | 32K | ~26 | ✓ | Toggle |
| DeepSeek R1 1.5B | DeepSeek | 1.2GB | 4GB | 8K | ~25 | — | Always on |
| Llama 3.2 3B | Meta | 2.1GB | 6GB | 8K | ~20 | — | — |
| Ministral 3 3B | Mistral AI | 2.7GB | 6GB | 32K | ~18 | — | — |
| Phi-4 Mini | Microsoft | 2.5GB | 8GB | 8K | ~16 | — | — |
| Qwen 3.5 4B | Alibaba | 2.9GB | 10GB | 32K | ~16 | ✓ | Toggle |
| Qwen 3.5 9B | Alibaba | 5.6GB | 12GB | 32K | ~10 | ✓ | Toggle |
★ Recommended default.
The models, by what they’re for
Qwen 3.5 0.8B — the recommended default
622MB · 5GB RAM · 32K context · ~40 tok/s · vision · thinking toggle
This is the model most people should start with, and the reason is memory headroom rather than raw capability. It supports vision, thinking mode, web search, and 30+ languages while keeping the smallest footprint of the multimodal family — which means a heavy turn combining web search and thinking stays well clear of the memory limit even on 8GB devices.
It is not the most capable model here. Visual benchmark scores are clearly below the 4B and 9B, and it will lose to larger models on multi-step reasoning. But it’s fast, it’s genuinely multimodal, and it runs everywhere. Hugging Face.
Qwen 3.5 4B — best all-around, if you have the memory
2.9GB · 10GB RAM · 32K context · ~16 tok/s · vision · thinking toggle
The strongest general model in the catalogue: exceptional reasoning, coding, and multilingual quality, with much better image understanding than the smaller Qwen 3.5 models.
The catch is the 10GB requirement, which rules out every 8GB iPhone. In practice this means a 12GB iPhone or an M-series iPad. If your device qualifies, this is the one to use for demanding work.
Qwen 3.5 2B — the practical multimodal middle
1.6GB · 6GB RAM · 32K context · ~26 tok/s · vision · thinking toggle
The smallest tier where image understanding is genuinely good for everyday photos. Stronger reasoning and coding than the 0.8B, at a larger memory cost. If you have 6GB or more and you use vision often, this is the upgrade that matters most.
Qwen 3.5 9B — the flagship
5.6GB · 12GB RAM · 32K context · ~10 tok/s · vision · thinking toggle
The best quality available on-device, and the most demanding by a distance. At around 10 tokens per second on an iPhone it’s usable but deliberate; on an M-series iPad Pro it roughly doubles, which is where this model belongs.
Worth it if you have the hardware and want the ceiling. Not worth it as a daily driver on a phone.
Phi-4 Mini — coding and maths
2.5GB · 8GB RAM · 8K context · ~16 tok/s
Microsoft trained Phi-4 Mini on curated synthetic data — maths problems, code, structured reasoning — rather than scraped web text, and it shows. On technical tasks it competes with models substantially larger.
The trade-offs are a short 8K context and weaker performance on open-ended creative work and non-English languages. For code review, debugging, and algorithm explanation it’s the sharpest tool here. Hugging Face.
Ministral 3 3B — efficient general purpose
2.7GB · 6GB RAM · 32K context · ~18 tok/s
Mistral AI’s edge-optimized model, with a full 32K context at a 6GB requirement — the best context-to-memory ratio among the non-Qwen models. Solid for chat, coding, and multilingual work. Mistral AI on Hugging Face.
Llama 3.2 3B — the reliable generalist
2.1GB · 6GB RAM · 8K context · ~20 tok/s
The most widely tested local model in the list, with a large ecosystem of tools and fine-tunes behind it. It behaves predictably and follows system prompts reliably, which matters if you have workflows already tuned for it. The 8K context is the main limitation. Hugging Face.
Llama 3.2 1B — light and dependable
800MB · 4GB RAM · 8K context · ~28 tok/s
A good lightweight generalist for 4GB devices where the multimodal models aren’t available.
DeepSeek R1 1.5B — visible reasoning in 1.2GB
1.2GB · 4GB RAM · 8K context · ~25 tok/s · always-on reasoning
DeepSeek’s R1 series popularized visible chain-of-thought, and this distillation brings it to a model that runs on 4GB. Unlike the Qwen models, its reasoning is always on rather than a toggle — so it’s a specialist for step-by-step problems rather than a general chat model. Hugging Face.
Qwen 3 1.7B and Qwen 3 0.6B — the lightweight thinkers
914MB / 317MB · 4GB RAM · 32K context · ~30 / ~45 tok/s · thinking toggle
Both offer a full 32K context and a thinking toggle at a 4GB requirement, which makes them the most capable options for older devices. Qwen 3 1.7B is the better all-rounder; Qwen 3 0.6B is the lightest model in the catalogue and the fastest.
Neither supports vision — that’s the trade for running on 4GB.
Which one should you download?
Start with Qwen 3.5 0.8B. It’s the default for good reason: multimodal, fast, and safe on memory. Most people never need more.
Move to Qwen 3.5 2B if you have 6GB or more and you use image input regularly.
Move to Qwen 3.5 4B if you have 10GB or more and want the best general quality available.
Add Phi-4 Mini if you write code and have 8GB.
Use Qwen 3 1.7B or 0.6B on 4GB devices, or when you want maximum speed.
Use Qwen 3.5 9B on an M-series iPad, where it runs properly.
One practical advantage of Cloaked: switching is instant. Keep a small fast model for everyday questions and a larger one for harder work — both stay on the device, and neither touches the cloud. For an honest read on where local models still fall short, see Is Local AI Good Enough Yet?.
Download Cloaked on the App Store to run any of these models on your iPhone — no account, no subscription, no data leaving your device.
Frequently asked questions
What is the best LLM model for iPhone?
Qwen 3.5 0.8B is the recommended default for most people — it is only 622MB, requires 5GB of RAM, and still supports vision, thinking mode, and 30+ languages. If your device has 10GB or more, Qwen 3.5 4B is the best all-around model in the catalogue.
Why does a 2.9GB model need 10GB of RAM?
The weights are only part of the requirement. The key-value cache grows with conversation length, the vision encoder and image tensors add roughly 0.5–1GB at peak, and iOS reserves a substantial share of memory for the system. The stated minimum has to cover the peak of a heavy turn.
What is the smallest LLM that still works well?
Qwen 3 0.6B at 317MB is the lightest option and runs on any supported device with 4GB of RAM. It handles simple Q&A, summarization, and text editing well. For vision and stronger reasoning in a still-small package, Qwen 3.5 0.8B at 622MB is the better choice.
Which models can read images?
The four Qwen 3.5 models — 0.8B, 2B, 4B, and 9B. They are natively multimodal, so image input uses the same weights as text. Larger models produce noticeably better image understanding, but even the 0.8B handles signs, menus, and screenshots well.
Does model size affect battery life?
Yes. Larger models require more memory bandwidth per token and sustain higher power draw. Generation is bursty rather than continuous, so a typical conversation costs far less than an hour of video playback, but a long session with a 9B model is noticeably more demanding than the same session with the 0.8B.