I think you’re looking one row too high, and that’s our fault for how the pane is laid out rather than yours.
There are two model settings in Settings ▸ AI, stacked one above the other:
Model — the one you’ve found. This is for writing: answers, summaries, drafting. It’s Ollama’s, and it’s the one that goes blank when Ollama isn’t running.
Search by meaning — the row just below it. That’s where MojoPad’s own model lives, and it’s the only place it appears.
Open that second dropdown and you should see three things: Automatic, MojoPad’s own — runs on this Mac, and then your installed models grouped underneath. Pick the middle one and you’re using ours.
Two things I got wrong, both of which sent you looking in the wrong place.
The release note is misleading if you already run Ollama. I wrote it as though MojoPad’s own model is what you get and your own are the alternative. For you it’s the other way round: with your models installed, Automatic uses yours, and MojoPad’s own is the fallback for a Mac that has none. So nothing changed on your setup when you updated, and nothing was going to. If you want to try ours you have to choose it by name. I’ve reworded that note.
And the label on the first row is just called “Model”, which is precisely what you’d click on when your question is “how do I choose the model”. It doesn’t say what it’s the model for. I’ve changed it to say “Writing & answers”, so the two rows read as the two different jobs they are.
There’s a third thing you probably saw and didn’t mention: with Ollama stopped, that first row also says “Ollama is not running, so AI features are off.” That sentence is now wrong — searching by meaning works perfectly well with Ollama off, which is the entire point of shipping a model. It’s fixed too, and thank you for walking into it.
One last thing, in case you do switch. Changing the Search by meaning setting rebuilds every wiki’s meaning index from scratch. That’s deliberate: two models don’t produce comparable numbers, and measuring one against the other doesn’t fail cleanly — it returns a confident score that means nothing. So MojoPad records which model built each index and refuses to mix them. It happens in the background, but on a big wiki you’ll notice it working.
And honestly — you may well want to stay on yours. Your embedding model is almost certainly larger than the one we ship, and larger generally reads your writing more finely. Ours exists so that somebody without Ollama isn’t left with nothing at all, which is what they got before.
Mark
