Models
Download, import and switch AI models, and get the best speed from your device.
Lumen runs open AI models in the GGUF format with llama.cpp, plus Apple Intelligence where it's available. Everything runs on your device.
Download a model
Open Models from the sidebar (More › Models).
- Featured models are tested and sized for phones: Qwen3, Llama 3.2, Gemma 3, Phi, DeepSeek R1 Distill, SmolLM, coder models for code mode and vision models that can see photos.
- Each model shows its size and how well it fits your device: Runs well, or a warning when it would be tight on memory.
- Tap Download. Downloads continue in the background and resume if the connection drops. Lumen checks each file's fingerprint (SHA-256) before using it.
- Pause a download from Models › Downloading and tap Resume later. It continues where it stopped, even after you close Lumen.
Find more models
- Hugging Face: search for any GGUF model, filter the results, pick a quantization and download it.
- Ollama library: search the Ollama registry and download models from there.
- Import: use Import .gguf file to pick a file from Files, or drop
.gguffiles into Files › On My iPhone › Lumen. Lumen adds them automatically.
Apple Intelligence
On supported devices with Apple Intelligence turned on, Apple Intelligence appears as a built-in model. It needs no download and runs on device through Apple's frameworks. Its context window is smaller than most downloaded models, so for long documents or code a downloaded model works better.
Choose which model answers
- Default model: in Models, set the model new chats use. Code mode can have its own default.
- Per chat: tap the model name in the composer to switch for that chat.
- Per assistant: custom assistants can have a preferred model (see Customize).
Speed and memory
Settings › Models & generation has the advanced controls:
- Context length: how much of the conversation the model remembers. Longer uses more memory; Lumen lowers it automatically if your device can't fit it.
- Max response: a cap on reply length.
- GPU acceleration: uses the device's GPU (Metal). Leave it on unless a model misbehaves.
- Compact memory: stores the model's working memory in a smaller format, so bigger models and longer chats fit.
- Free memory in background: unloads the model a little after you leave Lumen, so iOS keeps Lumen alive.
- Temperature, Top P, Top K, Min P, repeat penalty: fine-tune how creative or focused answers are.
Use Benchmark on a model to measure its speed on your device in tokens per second.
Delete a model
Swipe a model in Models, or long-press it, and choose Delete. Your chats stay; only the model file is removed.