Local models
Pick the model that reads your day — Qwen, Phi, DeepSeek and more — and switch it whenever. The model changes; your memory does not.
The model that answers your questions runs on your machine, and you choose which one. Nothing about that choice is permanent: switch tiers on a Tuesday, switch back on a Thursday, and your memory is untouched either way.
Models are managed in Settings → Local models.
Pick a size
The page offers three chat tiers, so the first decision is one you can make without knowing model names.
| Tier | Roughly | Good for |
|---|---|---|
| Small (0.5B) | ~500 MB | Any machine. Basic chat and short summaries. |
| Medium (7B) | ~5 GB | 16 GB RAM and up. Everyday questions and multi-turn chat. |
| Large (14B) | ~9 GB | A strong machine, 32 GB RAM. Best local quality. |
Each tier lists the hardware variants available for it — NPU, GPU, CPU — so a Copilot+ PC can run a model on its neural engine while a laptop with a discrete card runs the same size on the GPU. Pick a variant and LUCI downloads it and makes it the default.
Or pick the exact model
More models opens the full catalog: search by name, filter by CPU / GPU / NPU, and switch between what you have downloaded and what is available. This is where Phi, DeepSeek, and the rest of the catalog live, next to the Qwen builds the tiers use.
The ? next to the chat model opens a hardware guide — a table of machine tiers with the models that fit each one, with separate tables for Windows and Apple Silicon, because they have different constraints. Windows machines are ranked by VRAM; Macs by total unified memory, since weights compete with macOS, your browser, and LUCI itself.
Switching
Setting a model as default points both the chat and the daily summaries at it.
On Windows the new default is loaded into memory for you. On a Mac, press Load on the variant you want — loading is explicit there because the runtime does not reliably hand memory back on unload, so LUCI does not load models behind your back.
Model execution on Windows runs through Microsoft Foundry Local. Its docs are the reference for which hardware variants exist and what each one needs.
Downloaded models live in LUCI's own folder in your home directory, not in a system cache. You can delete any of them from the same page; deleting a model does not delete anything it produced.
Speech
Transcription uses local Whisper models, sized the same way — tiny and base on modest machines, up to large-v3-turbo on a strong one. The hardware guide lists the recommended one per tier alongside the chat models, with rough speeds.
Offline
Downloading a model needs a connection, once. After that, chat, summaries, transcription, and search all run with the network off.
Related
- System requirements for the machine LUCI expects
- Search and insights for what search does with the index
- Privacy on device for what stays here
- Microsoft Foundry Local — Microsoft's own docs for the Windows runtime, including the source on GitHub
Was this page helpful?