Apple · A19 Pro · Phone
Best AI Models for iPhone Air (12 GB)
Runs 1B models well; dense 3–4B fits but is slow at 15 tok/s, and 8–9B only just fit in the 7 GB iOS gives one app.
| # | Model | Speed | llmrun Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Qwen3.5 4BAlibabaHow to run → Q4_K_M · 3.4 GB · 14.6 t/s | 14.6 tok/s | Q4_K_M · 3.4 GB | |
| 2 | Gemma 3 4B ITGoogle Q4_K_M · 5.0 GB · 16.8 t/s | 16.8 tok/s | Q4_K_M · 5.0 GB | |
| 3 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 41.0 t/s | 41.0 tok/s | Q4_K_M · 5.5 GB | |
| 4 | Qwen3.5 2BAlibaba Q4_K_M · 1.9 GB · 27.0 t/s | 27.0 tok/s | Q4_K_M · 1.9 GB | |
| 5 | MiniCPM5 2BOpenBMB Q4_K_M · 2.5 GB · 25.1 t/s | 25.1 tok/s | Q4_K_M · 2.5 GB |
| # | Model | Speed | Reasoning Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Q4_K_M · 5.0 GB · 16.8 t/s | 16.8 tok/s | Q4_K_M · 5.0 GB | |
| 2 | Ling 3.0 TinyInclusion AI Q4_K_M · 5.5 GB · 41.0 t/s | 41.0 tok/s | Q4_K_M · 5.5 GB | |
| 3 | Qwen3.5 2BAlibaba Q4_K_M · 1.9 GB · 27.0 t/s | 27.0 tok/s | Q4_K_M · 1.9 GB | |
| 4 | MiniCPM5 2BOpenBMB Q4_K_M · 2.5 GB · 25.1 t/s | 25.1 tok/s | Q4_K_M · 2.5 GB | |
| 5 | Spark X2.6 4BXHToken Q4_K_M · 4.3 GB · 16.1 t/s | 16.1 tok/s | Q4_K_M · 4.3 GB |
| # | Model | Speed | Vision Score | Quant · Memory |
|---|---|---|---|---|
| 1 | Q4_K_M · 5.0 GB · 16.8 t/s | 16.8 tok/s | Q4_K_M · 5.0 GB | |
| 2 | Qwen3.5 4BAlibaba Q4_K_M · 3.4 GB · 14.6 t/s | 14.6 tok/s | Q4_K_M · 3.4 GB | |
| 3 | Qwen3.5 2BAlibaba Q4_K_M · 1.9 GB · 27.0 t/s | 27.0 tok/s | Q4_K_M · 1.9 GB | |
| 4 | MiniCPM V 4.7 1BOpenBMB Q4_K_M · 1.6 GB · 37.3 t/s | 37.3 tok/s | Q4_K_M · 1.6 GB | |
| 5 | LFM3 Omni 2BLiquid AI Q4_K_M · 2.4 GB · 28.1 t/s | 28.1 tok/s | Q4_K_M · 2.4 GB |
The iPhone Air suits smaller models: it runs 1B models well. iOS lets one app use about 7 GB of the 12 GB, so 8–9B models barely fit. Prompt reading is its weak spot: a 4K-token prompt waits 27 s before the first word on Qwen3.5 4B. A19 Pro chip, released 2025.
iPhone Air strengths and weaknesses
Strengths
- Runs 1B models: Llama 3.2 1B Instruct at 58.2 tok/s in 2.3 GB
Weaknesses
- 14B models don't fit: Qwen3 14B needs 11.9 GB at Q4 vs 7 GB usable here
- Little room for context: Qwen3.5 9B uses 6.6 of 7 GB usable, 0.4 GB left
- Slow prompt reading: A 32K-token prompt on Qwen3.5 4B waits 3.5 min for the first token
Memory for models
7 GB of 12 GB14B models need 11.9 GB at Q4, more than the 7 GB available.
How much memory iOS lets one app use decides what phone works for local AI, not the camera or the screen.
Phones run small models. Output speed follows memory bandwidth (68.2 GB/s); grey rows need more than the 7 GB iOS gives one app.
- 1BLlama 3.2 1B Instruct · Q4_K_M · 2.3 GB58.2tok/s
- 3–4BQwen3.5 4B · Q4_K_M · 3.4 GB14.6tok/s
- 7.5tok/s
- 14BQwen3 14B · not at 4-bitneeds 11.9 GB
Estimated speeds: Q4_K_M, 4K context, one user. How we estimate
iPhone generations: memory one app can use
A17 Pro → A19 Pro- iPhone 15 ProA17 Pro · 20234.5 GBup to 4B
- iPhone 16 Pro MaxA18 Pro · 20244.5 GBup to 4B
- iPhone 17A19 · 20255 GBup to 4B
- iPhone AirA19 Pro · 20257 GBup to 9B (tight)
Every phone here has 8 GB or more, but iOS caps one app well below that. The iPhone Air is the first to fit an 8–9B model.
All 783 models on the iPhone Air
Featured onlyBest quant that fits in 7 GB, ordered by llmrun Score. Fit grades how well it runs here, S best. How ranking works
| Model | Quant | VRAM | Speed | Context | Score | Fit |
|---|---|---|---|---|---|---|
North Micro Vision InstructVisionChat | Q4_K_M | 3.7 GB | 23.5 t/s | 500K | B64, Decent | |
JiRackUltra 1BChatFunctions | Q4_K_M | 1.8 GB | 33.6 t/s | 131K | A72, Runs well | |
D1 3BVisionFunctions | Q4_K_M | 2.6 GB | 20.8 t/s | 33K | B61, Decent | |
MinerU2.5 Pro 2604 1.2BVision | Q4_K_M | 1.1 GB | 46.8 t/s | 8K | A78, Runs well | |
| BF16 | 0.4 GB | 111.0 t/s | 1K | S93, Runs great | ||
Pythia 14MChat | Q4_K_M | 0.0 GB | 4774.0 t/s | 2K | S100, Runs great | |
WeVisOne VL 2BVisionFunctions | Q4_K_M | 3.6 GB | 23.9 t/s | 262K | B64, Decent | |
| Q4_K_M | 0.5 GB | 159.1 t/s | 66K | S97, Runs great | ||
Tiny RemoteForCausalLMChatFunctions | BF16 | 0.5 GB | 159.1 t/s | 131K | S97, Runs great | |
Ornith 1.0 9BChatFunctions | Q4_K_M | 6.5 GB | 7.7 t/s | 262K | D26, Barely runs | |
| Q4_K_M | 2.6 GB | 48.2 t/s | — | A79, Runs well | ||
Tiny Qwen3ForCausalLM Instruct 2507ChatFunctions | Q4_K_M | 0.5 GB | 159.1 t/s | 262K | S97, Runs great | |
Granite Vision 4.1 4BVisionFunctions | Q4_K_M | 4.0 GB | 16.6 t/s | 131K | B56, Decent | |
FrogNano 4B 2609ChatFunctions | Q4_K_M | 3.4 GB | 14.6 t/s | 262K | B54, Decent | |
WeVisOne VL 4BVisionFunctions | Q4_K_M | 4.7 GB | 14.1 t/s | 262K | B53, Decent |
Specs
- Chip
- A19 Pro
- CPU cores
- 6
- GPU cores
- 5
- Unified memory
- 12 GB LPDDR5X-8533
- For models
- 7 GB (iOS keeps 5 GB)
- Memory bandwidth
- 68.2 GB/s
- FP16 (dense)
- 7.2 TFLOPS
- NPU
- 16-core Neural Engine
- Released
- 2025-09-19
iPhone Air FAQ
Can the iPhone Air run an LLM offline?
Yes, small ones. Qwen3.5 4B runs at about 14.6 tok/s, and 1B models run around 58 tok/s. iOS lets one app use about 7 GB of the 12 GB, so larger models don't fit.
How much memory is available for AI on the iPhone Air?
About 7 GB of the 12 GB. iOS keeps roughly 5 GB for itself and other apps.
What is the largest model the iPhone Air can run?
Qwen3.5 9B (9.7B) is the largest at 4-bit or better: at Q4_K_M it uses 6.6 GB, 95% of the available memory, and runs at about 7.5 tok/s.
How long does the iPhone Air take to read a long prompt?
On Qwen3.5 4B, about 27 s for a 4K-token prompt and 3.5 min for a 32K-token document before the first word appears.