Is NilaMind a therapist / can it replace one? No. It's an experimental self-help companion — explicitly not a medical device, therapy, or a crisis service. It doesn't diagnose or treat anything. It's designed to complement real human support and care, and it actively nudges you back toward people (see the dependency guard).
Does anything I say leave my phone? No. Replies are generated by a model running locally; there's no backend, no account, and no analytics. The only network activity in normal use is the one-time model download on first run. The only way personal content can leave is if you explicitly donate a single, PII-scrubbed example to help improve the model — opt-in and revocable. See Privacy & Data.
Why is the first reply so slow? The model processes a large system prompt once per session ("prefill"). Subsequent replies reuse that work and are much faster. It's a deliberate trade — a richer 4B model that's slower to warm up, over a smaller, faster, weaker one. See The On-Device Brain.
Why isn't the model included in the app? GGUF models are large and licensed separately, and a crisis-capable app shouldn't ship a brain by default. The app downloads it on first run (verified) or you can side-load your own. See Model Provisioning.
Can I use a different model?
Yes — drop in any GGUF, or wire a different on-device backend via the localLlm seam. Re-test the §9 safety layer against your model first.
Why isn't it on the main F-Droid store? Because it relies on a prebuilt llama.cpp inference binary, and f-droid.org builds everything from source. The FOSS-ecosystem route is IzzyOnDroid. See Distribution.
Which devices are supported? Android 7.0+ (API 24), arm64-v8a, with enough free storage (~1.5 GB) and RAM for the model. The default brain (Qwen2.5-1.5B, ~1.1 GB) runs on most recent phones; only the optional side-loaded 4B model strains low-RAM devices.
Is there an iOS version? Not currently. The codebase is React/Capacitor, so iOS is conceivable, but the on-device LLM runtime and the build are Android-only today.
What data does it collect about me? None, by design. See Privacy & Data.
I found a safety gap / the crisis layer missed something. Please treat that seriously and report it. The crisis layer is high-recall but imperfect — see Contributing. And remember: software is not a reliable way to reach help in an emergency.
Is it "done" / production-ready? No. It's experimental and personal — not clinically validated, no support guarantees. Use at your own risk.
Who built it, and why? A personal research project by someone with lived experience of mental illness, to support people who are struggling and may not feel able to open up. Shared openly under the rule: help is the only metric — never gather data at any cost.