Castellan
Reference · Castellan
Which AI models the staff use
The small AI models Castellan sets up on this PC for chat, pictures and search, which one each kind of NPU gets, and where they come from.
- Article
- 1110
- Applies to
- Castellan 0.17.1
- Last reviewed
- For
- For everyone
The models#
The staff use small, open AI models that run on this PC, never online. Castellan's setup picks them for this PC's NPU, graphics card and processor, and the Smith keeps them running. A model only describes and answers: code still decides everything that matters.
| What for | Model |
|---|---|
| Chat and pictures: the Butler's answers, and the staff's labels, summaries and descriptions of screenshots | One model for both, picked for this PC (below) |
| Search: finding passages and files by what they mean | Qwen3-Embedding-0.6B |
| Heiward's duplicate photos and videos | DINOv2 |
| Reading the words in pictures and PDFs | Windows' own text recognition, on the processor |
Which model this PC gets#
| On | Chat and pictures |
|---|---|
| A Snapdragon X2 Elite or X2 Plus NPU | Qwen3-VL-2B |
| A Snapdragon X, X Plus or X Elite NPU | Qwen3-VL-4B |
| An AMD Ryzen AI 300 or later NPU | Qwen3.5-2B |
| An Intel Core Ultra NPU | Qwen3-4B-Instruct-2507 for chat; pictures go to the graphics or the processor |
| A graphics card, or the processor | Qwen3-VL-2B |
Castellan uses the smallest model that does the job well. The 2B models passed the same checks as the 4B ones, wrote about twice as fast and used less of the PC's memory, so a bigger model is used only where a test shows it's needed.
Where they come from#
Setup downloads the models once, from Hugging Face or through the NPU's own program, and checks each one before it's used. Castellan's updates can bring newer models, and the Smith tries a new one before it trusts it. The Qwen models are made by Alibaba's Qwen team, and DINOv2 by Meta, all under the Apache 2.0 license.
Related articles
Is this page right?
If something on it is wrong or out of date, tell us and we'll fix the page.
Still stuck? Write to support@castellan-software.com and mention article 1110. Every version of Castellan, and what changed in it, is in its release notes.