A 0.5B-parameter model is, by 2026 standards, embarrassing. It's worse at math than a calculator. It hallucinates more than a sleep-deprived intern. Its idea of "deep reflection" can devolve into Hallmark-card pap if you let it.
And it's still the right default for Graphophone. Here's why.
It works offline, today, on a four-year-old phone.
"It works" is a higher bar than "it's clever." If reflection is something you want to do at 2 AM in airplane mode, in a cabin without signal, on a plane next to your spouse — then the model that's there is infinitely better than the model that's smart but elsewhere.
It removes "should I" from every interaction.
When the smart option requires sending your entry to a cloud API, you have to decide, every time, whether this entry is private enough to keep on-device. That friction kills the journaling habit. With the bundled model, there is no decision: nothing leaves the device. The smarter cloud option is opt-in per surface, so you only think about it on the surfaces where it pays off.
It punishes prompt-tinkering.
A small model rewards prompts that ground the work in *your* writing rather than asking the model to be clever in general. The pattern detective, the summary surfaces, the reflection chat — they all work because the system prompt is concise and the retrieval is precise. The model is a synthesizer, not an oracle. Building against that constraint makes the surfaces feel less like Twitter and more like the journal itself.
It buys us the upgrade path on our own terms.
Apple Foundation Models and Gemini Nano will slot in as Tier 1 on the devices that have one. Downloadable larger open-weight models — the ones whose licences let us ship them commercially — slot in as Tier 3. Your home server and your cloud key slot in as Tiers 4 and 5. The bundled model is the floor under all of that — and the floor is the part that actually ships on day one.
Correction (2026-06-13): the default on-device model is Qwen 2.5 0.5B (~0.5B parameters), not 1.5B, and it is downloaded once on first run rather than bundled inside the app binary. It runs via the candle pure-Rust engine. The argument above is unchanged; the numbers and the "ships with the app" framing are corrected here.
