Software the app is built on
CuriousLM runs models with llama.cpp (MIT License): through wllama (MIT License) in the browser, and compiled into the Android app together with Arm KleidiAI (Apache License 2.0). Android also uses Capacitor (MIT License) and AndroidX components (Apache License 2.0). The web app uses React, Next.js, Dexie, pdf.js, Mammoth, KaTeX and highlight.js under their respective open-source licences. Voice runs on Transformers.js (Apache License 2.0) in the browser and ONNX Runtime (MIT License) on Android.
Models
Qwen, Gemma and MiniCPM models are distributed under Apache License 2.0, as shown in the model library. LFM models from Liquid AI use the LFM Open License 1.0. Its commercial grant is conditional on the relevant entity remaining below US$10 million annual revenue. A free application does not remove that condition.
Curious 2B is CuriousLM's own model: Qwen3.5 2B (Copyright 2026 Alibaba Cloud), fine-tuned by CuriousLM for this app. Releases up to 261018 are distributed under Apache License 2.0. From 261019 on, Curious 2B also learns from datasets shared for non-commercial use, so those releases are distributed under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 licence. Its training data includes MASSIVE by Amazon and When2Call by NVIDIA, both shared under the Creative Commons Attribution 4.0 licence.
Voices
Speech is turned into text by Moonshine (MIT License) or Whisper (MIT License), and read aloud by Kokoro 82M (Apache License 2.0), all running on your device.
Pocket Wikipedia
Pocket Wikipedia holds the opening sections of Wikipedia articles, written by Wikipedia's contributors and shared under the Creative Commons Attribution-ShareAlike 4.0 licence. CuriousLM shortens the articles to their opening sections and names Wikipedia under every answer that used it, with a link to the article.
How models are delivered
The catalog records each model's exact Hugging Face repository commit, file name, byte count, SHA-256 and licence. Files download directly from Hugging Face to your device only when you ask, and a file whose size or checksum does not match is discarded. CuriousLM publishes one model itself, Curious 2B, on its own Hugging Face page, and does not host or modify any other.