What does the Think button do?
A normal reply starts straight away: the model writes the answer word by word as it goes. With Think on, the model first writes notes to itself, working through the question, and only then writes the answer. While it thinks, Nashi shows Thinking…; the notes stay hidden, and you see the finished answer.
That extra step matters most for small models. A model that answers instantly can rush a sum or skip a step. Giving it room to reason catches many of those mistakes.
When should I turn Think on?
Turn it on for questions that have steps:
- Maths and timing. "My train leaves at 9:40 and takes 2 h 55 min. Then I have a 25 min wait and a 1 h 10 min bus. When do I arrive?" (14:10.)
- Planning. A day's schedule, splitting a bill, a packing list with constraints.
- Logic and code. A puzzle, a tricky spreadsheet formula, a bug in a few lines of code.
Leave it off for chat, rewriting, quick facts, ideas and translation. Those don't need reasoning, and you get the answer sooner.
Which models can think?
| Model | Think | Starts |
|---|---|---|
| LFM3 1.7B | Yes | On |
| Granite 4.2 3B | Yes | Off |
| Gemma 4 E2B | Yes | Off |
| Gemma 4 E4B | Yes | Off |
| LFM2.5 1.2B | No | No button |
LFM3 was built to reason, so it starts with Think on. You can switch it off, and Nashi remembers your choice for each model. See choosing a model in Nashi for how the models compare.
How long does thinking take?
Seconds, usually. Nashi limits thinking to 512 tokens, roughly 380 words, then the model has to answer, so it never thinks for minutes. On an iPhone 15 Pro Max, LFM3 explained the water cycle in about 120 words in under 18 seconds, thinking included. Bigger models and older phones take longer.
Does thinking happen on my phone?
Yes. Thinking is just more text the model writes, and the model runs on your device. Nothing is sent anywhere, and Think works in Airplane Mode like the rest of Nashi. More on that in how offline AI works on iPhone.