Models.
What runs on the phone, and under which licence.

01Default catalog

Files: Qwen3 0.6B (Q4_0) and Qwen3 1.7B (Q4_K_M) from ggml-org, SmolLM2 1.7B (Q4_K_M) from HuggingFaceTB, SmolVLM 500M (Q8_0 + projector) from ggml-org, Whisper base (multilingual, q5_1) from the whisper.cpp model repository. Every file is checked against its sha256 after download.

The tier only matters while the paywall is on (Purchases); while it is off every model is available.

02Kinds

• llm: chat models. The user picks one in Models; it is the one in use.
• vision: answers about photos; needs its projector file (mmprojUrl), downloaded together.
• whisper: voice dictation (whisper.cpp ggml-*.bin files).

03Add a model

1. Check the licence. Only add models whose licence allows commercial use without use restrictions (Apache-2.0 and MIT are safe). Avoid licences with acceptable-use policies or non-commercial terms.
2. Pick the GGUF file on Hugging Face and copy its direct link: https://huggingface.co/<repo>/resolve/main/<file>.gguf.
3. Get the size and sha256 from the file page on Hugging Face (it shows both), or after downloading it: shasum -a 256 file.gguf and stat -f %z file.gguf.
4. Add one line in the console (Model catalog), or to models in app/assets/config/default.json (then run python3 tool/gen_default_config.py):

   {"id":"qwen3-4b","name":"Qwen3 4B","kind":"llm","url":"https://huggingface.co/…/Qwen3-4B-Q4_K_M.gguf","sha256":"…","sizeBytes":<exact size>,"minRamGb":8,"tier":"pro","license":"Apache-2.0","licenseUrl":"https://huggingface.co/Qwen/Qwen3-4B","nCtx":4096,"systemSuffix":" /no_think","default":false}

A Qwen3 model needs "systemSuffix":" /no_think" to answer directly; the app also hides any <think> block a model writes.