How fast does Qwen2-VL 7B (Vision) run on iPhone?
On an iPhone Air (A19, iOS 26.6.1), Qwen2-VL 7B (Vision) generated 11.6 tokens per second in Pocket AI Lab 1.1.0, averaged over 3 runs of 200 tokens (± 4.8).
On the standard llama-bench style probe it read a 512-token prompt at about 46 tok/s and generated 128 tokens at 4.8 tok/s. Cold load took 2.5 s and the process peaked at 5,461 MB. The phone heated up during this run and decode speed dropped 59% between repetitions, so on a cool phone it can be faster.
How much memory and storage does Qwen2-VL 7B (Vision) need?
The catalog budgets 5.2 GB of RAM and 4.7 GB of storage for the 4bit build of Qwen2-VL 7B (Vision). The measured peak footprint on iPhone Air was 5,461 MB.
Its backend copies weights into the app's memory, so the app checks it against the per-app limit iOS enforces. On a 6 GB iPhone that limit is much lower than 6 GB.
Which iPhones can run Qwen2-VL 7B (Vision)?
Qwen2-VL 7B (Vision) runs on iPhones with 8 GB of RAM or more. On 8 GB and 12 GB phones it fits with little headroom, so long chats may make it reload.
| RAM | iPhones | Pocket AI Lab |
|---|---|---|
| 3–4 GB | iPhone 11, 11 Pro, 12, 12 mini, 13, 13 mini, SE (2nd and 3rd gen) | Not measured yet |
| 6 GB | iPhone 12 Pro, 13 Pro, 14, 14 Plus, 14 Pro, 15, 15 Plus | ✗ Too big |
| 8 GB | iPhone 15 Pro, 16, 16 Plus, 16 Pro, 16e, 17, 17e | ≈ Runs, little headroom |
| 12 GB | iPhone 17 Pro, 17 Pro Max, iPhone Air | ≈ Runs, little headroom |
Based on the memory budgets the app measures on each class of iPhone (October 2026). Before downloading, the app checks your exact phone and only offers what fits.
What can Qwen2-VL 7B (Vision) do in Pocket AI Lab?
It's a vision-language model, so besides text chat it can describe photos and frames sampled from a video you attach to a message.
How did Qwen2-VL 7B (Vision) score on the quality suite?
It passed 75 of 118 tasks, a score of 63.6 out of 100 on suite v3.
Tasks are graded by deterministic checks (exact numbers, regex, word counts, valid JSON), with greedy decoding. See the method.
All runs of Qwen2-VL 7B (Vision)
| Model | Gen tok/s | pp512 / tg128 | Load s | Peak MB | Score | Date |
|---|---|---|---|---|---|---|
| Qwen2-VL 7B (Vision) | 11.5 ± 3.9 | ~46 / 5.3 | 3.5 | 5,464 | 78.3 | 2026-09-03 |
| Qwen2-VL 7B (Vision) | 11.6 ± 4.8 | ~46 / 4.8 | 2.5 | 5,461 | 63.6 | 2026-10-22 |
Other Qwen models in the catalog
- Qwen 2.5 3B, MLX, 4bit, 26.3 tok/s
- Qwen 2.5 14B, MLX, 4bit
- Qwen 3.5 0.8B (GGUF, Vision), llama.cpp, Q4_K_M, 73.7 tok/s
- Qwen 3.5 2B (GGUF, Vision), llama.cpp, Q4_K_M, 36.6 tok/s
- Qwen 3.5 4B (GGUF, Vision), llama.cpp, Q4_K_M, 14.7 tok/s (throttled)
- Qwen2.5 1.5B (Core ML), Core ML, int8-W8A16, 3.0 tok/s
How do I install Qwen2-VL 7B (Vision)?
Open the Models tab in Pocket AI Lab and tap Qwen2-VL 7B (Vision). The app checks it against your phone's memory, downloads 4.7 GB from Hugging Face and it's ready for offline chat.