Why does a local AI chat download a model before the first reply?
Author: SunClaw Team · Updated: 2026-11-17
The short answer
Because the weights that produce a reply are not inside the app install. Without that file, the phone has no model to run.
When the selected model is not on the device yet, PrivAI shows a download note: the internet is only needed for this first download, and after that it works offline. If the file is already there, Confirm loads it and does not download it again. The empty chat says nothing is uploaded. After one download you can keep a multi-turn chat offline.
Which model to pick is in which local model on iPhone. What offline means is in run a model locally on iPhone.
How big the recommended one is
The catalog marks Qwen3 1.7B as recommended. It is stored as 1000 MB and shown as 1.0 GB. Anything of 1000 MB or more is shown as MB divided by 1000, to one decimal, not divided by 1024. Its listed minimum RAM is 4 GB.
The phone's RAM is physical memory converted to GB and rounded to an integer. If that integer is below the model's listed minimum, the chat shows a warning. The warning is not a blocked download. Pick a smaller model and the warning is recomputed from that model's minimum.
Other current, non-legacy entries use the same display:
| Model | Stored size | On screen | Listed minimum RAM |
|---|---|---|---|
| Qwen3 0.6B | 400 MB | 400 MB | 3 GB |
| Qwen3.5 0.8B | 625 MB | 625 MB | 4 GB |
| Gemma 3 1B | 730 MB | 730 MB | 3 GB |
| Qwen3 1.7B (recommended) | 1000 MB | 1.0 GB | 4 GB |
| Qwen3.5 2B | 1700 MB | 1.7 GB | 6 GB |
| Llama 3.2 3B | 1800 MB | 1.8 GB | 6 GB |
| Qwen3 4B | 2300 MB | 2.3 GB | 6 GB |
| Qwen3.5 4B | 3000 MB | 3.0 GB | 8 GB |
Older entries stay in the catalog so a file you already downloaded can still be opened. A new download follows the entry you select now. Changing models is a different file. If that file is not local, it is another download.
After the file is local, the chat still has caps
The default temperature is 0.6. One reply stops at 2048 tokens, so a small model does not repeat without a limit. Context is counted in characters, with a budget of 8000. Past that, the oldest turns are dropped. The whole history is not kept in memory forever.
Those caps apply after a model is loaded. Until the file is on the device, there is no model to load and no first reply.
The confirm sheet names the download source. The default points at huggingface.co, and the fallback mirror is hf-mirror.com. The note for mainland China says to set the mirror under Settings, Mirror. The mirror only changes where the file comes from. It does not upload the chat.
FAQ
Q: Why is there another download of hundreds of MB, or several GB, after install? A: The install does not contain the weights. The first time, the selected model has to land on the device. The same file then loads offline.
Q: If I already downloaded it, will it download again? A: Not the same file. Confirm loads it. A different model that is not on the device yet is a new download.
Q: Does the RAM warning mean I cannot download? A: The warning means the phone's integer RAM is below that model's listed minimum. It is a warning. It is not the download control turned off.
Q: Is the chat uploaded? A: The empty state says nothing is uploaded. The network is for the first download of the model file.
Download the recommended one, then talk
If storage is tight, start from the recommended Qwen3 1.7B: 1.0 GB on screen, listed minimum 4 GB of RAM. Chat after it is local. For a smaller or larger file, use the table, and expect another download.