Local Model

Local models are built into Cherry Studio,and can run offline after download small models: they do not use any provider API and do not require an API Key. They are small in size, run on your own computer, and are meant to provide fallback support for basic capabilities that aren't worth configuring a separate cloud model for.

After opening Settings → Local Models to manage them:

Local models: ① two built-in local models — the local embedding model + the local OCR model (both shown as "Ready" in the image; you can click the delete icon on the right to remove them)

There are currently two built-in types of local models:

Local models	Base	Size	Purpose
local embedding model	Qwen3 Embedding 0.6B	About 614 MB	Converts text into vectors for Knowledge base retrieval, recall, and similar scenarios
Local OCR model	PaddleOCR PP-OCRv6	About 140 MB	Offline recognition of text in images/scans for OCR function calls

Download and status

• A status badge will appear next to the model name: for cards not downloaded at the bottom there is a full-width "Download" button; click it to start downloading. After the download completes, the badge changes to Ready.
• For models that are ready, you can click the Delete icon on the right to remove them and free up disk space; you can download them again when needed. (If the embedding model is still being used by the knowledge base, deletion will be refused and the weights will be kept.)
• When a few platforms/architectures do not support local inference, the panel will show "This platform does not support local models"; downloads are not available in that case.

If one mirror is unavailable during download, Cherry Studio will automatically try other download sources. After the download completes, the inference process of the local embedding model runs on your machine and does not require an internet connection.

If the page shows [The model files are incomplete, please re-download to repair.], it means the local cache is missing required files. Deleting or re-downloading the model will fix it; do not manually piece together model files.

Local models are optional. Model services If you have configured a cloud embedding model in it, or the system's built-in OCR is sufficient, you don't need to download them.

When to use local models

• No cloud embedding model / don't want to pay separately for the knowledge base: download the local embedding model, and the knowledge base can index and retrieve completely offline.
• Need offline OCR: in scenarios with no network, or when you don't want images uploaded to third parties, download the local OCR model and use it together with OCR settings to select "Local PaddleOCR."
• Privacy first: all computations are completed locally, and your content never leaves your computer.

Local models are a lightweight solution that is "good enough." If you have very high requirements for retrieval accuracy or recognition accuracy, cloud-based Embedding model, and more powerful OCR services are usually better.

---

Get help and submit feedback

If you encounter any questions, bugs, or have suggestions for feature improvements during configuration or use, please refer to Feedback and Suggestions for the official channels provided.