# A small AI that runs on your own PC

> Chat with Qwen3.5 2B running locally through llama.cpp. No account, no subscription, nothing you write leaves your PC. One-time 1.82 GB download; reads pictures too.

Part of [PerchDeck](https://perchdeck.app/) for Windows. Canonical page: https://perchdeck.app/features/ai-chat

A chat with a small AI model that runs on your PC, with no account and no subscription. Nothing you write leaves your PC.

## What it does
- **What runs.** Qwen3.5 2B (by Alibaba, Apache 2.0), 4-bit, with its vision part so it can look at pictures, running on your CPU through llama.cpp (MIT).
- **One-time download.** About 1.82 GB from Hugging Face, with progress, stop and resume, and a checksum check before use. Apart from the license check, it is the only thing in PerchDeck that uses the internet. After that, the chat itself needs no connection.
- **Chat, with pictures.** Replies appear word by word as formatted text, with Copy on code blocks. Attach a PNG, JPG, WebP, GIF or BMP up to 25 MB, or paste a screenshot with Ctrl+V.
- **Your chats stay here.** Earlier chats are kept on this PC (turn it off in Settings). Pictures are never saved, only the text. You can delete one chat, all chats, or the model and all chats.

## Good to know
- Speed measured on an Intel Core i7-4800MQ from 2013 with no graphics card: first word in about 1 second and about 10 tokens a second for text. A large screenshot with small text took about 2 minutes before the first word. Newer PCs should be faster; that is not measured.
- It needs about 3 GB of free memory while it works (about 2.6–2.9 GB measured). A PC that reports 4 GB or less gets a warning, and the app recommends 8 GB or more. The model unloads itself after 10 minutes unused.
- If the download can't connect, the app says to check your internet connection, VPN or firewall and press Download again. Anything already downloaded is kept, so it carries on where it stopped.
- It is a 2-billion-parameter model. It makes mistakes and can be wrong with confidence: “AI can make mistakes. Check important information.” It is not comparable to large cloud models.
- The “Think” switch (off by default) makes it reason step by step first. It is slower, and on a PC without a graphics card may not answer within minutes.
- It only chats and reads pictures you attach. It does not act for you on files, to-dos or the web.

## Questions
### Does the AI send what I write to the cloud?
No. After the one-time model download, the chat talks to a helper on your own PC (127.0.0.1) that only PerchDeck can use.

### How big is the AI download and what do I need?
About 1.82 GB, once, and about 3 GB of free memory while it runs. A PC with 4 GB or less gets a warning, and 8 GB or more is recommended. The model is not part of the app install; it is a separate, optional download.

### The AI download won't start. What can I do?
Check your internet connection, VPN or firewall, then press Download again. PerchDeck tries both kinds of internet address, and anything already downloaded is kept, so it carries on where it stopped.

## Related
- [Note](https://perchdeck.app/features/note)
- [Clipboard history](https://perchdeck.app/features/clipboard-history)
- [Teleprompter](https://perchdeck.app/features/teleprompter)