Models and assignments
Curator runs every AI feature on local models through Ollama. Instead of one model doing everything, Curator assigns a model to each task. You can change any assignment in Settings under AI & Models.
The six tasks
Section titled “The six tasks”| Task | What it powers |
|---|---|
| Vision / Analysis | Reads images and scanned PDFs so their content becomes searchable. |
| OCR | Pulls text out of scanned documents and images. |
| Embeddings | Turns text into meaning, which powers search by meaning and the documents chat draws from. |
| Chat | Answers your questions and cites the files it used. |
| File Processing | Writes the AI description, category, and other metadata for each file. |
| Organization | Groups files into clusters on the Organize screen. |
Pulling models
Section titled “Pulling models”Models are downloaded, or pulled, through Ollama. Curator can pull its recommended models straight from Settings under AI & Models.
Recommended defaults
Section titled “Recommended defaults”These are the models Curator suggests out of the box. Vision and OCR share one model, and File Processing and Organization share another, so four downloads cover all six tasks.
| Task | Default model | Approx. download |
|---|---|---|
| Vision / Analysis | minicpm-v4.5 | 6.1 GB |
| OCR | minicpm-v4.5 (same as Vision) | shared |
| Embeddings | nomic-embed-text-v2-moe | 1.0 GB |
| Chat | qwen3.5:9b-mlx | 8.9 GB |
| File Processing | maternion/lfm2.5 | 5.2 GB |
| Organization | maternion/lfm2.5 (same as File Processing) | shared |
Sizes are the on-disk download and are approximate. Ollama confirms the exact size on the first pull.
Fitting your Mac’s memory
Section titled “Fitting your Mac’s memory”Onboarding reads how much memory your Mac has and warns before you pull a model that looks too large, using a rough rule of leaving about a third of your memory free for the system and other apps. The warning does not block the download, it just flags that the model may load slowly or run out of memory. Larger models are more capable but use more memory, and processing more files at once uses more still.
The Ollama connection
Section titled “The Ollama connection”Curator talks to Ollama at a single address, the Ollama Base URL. The default is http://localhost:11434, which is Ollama running on the same Mac. You can point it at Ollama on another computer on your network instead.
The connection test in Settings shows one of four states.
| Badge | Meaning |
|---|---|
| Not tested | You have not run the test yet. |
| Testing | The test is in progress. |
| Connected | Curator reached Ollama and AI features are available. |
| Offline | Curator could not reach Ollama. AI features stay off until it can. |
To understand what Ollama is and why Curator depends on it, read Ollama and models.