Updated November 17, 2026. This post first said PIM was moving its default from gemma4 to mistral-nemo. That didn’t last; here’s what PIM does now.
How we choose
General benchmarks didn’t tell us what we needed to know: whether a model gets PIM’s own requests right. So every candidate gets the same test, made from the requests PIM sends: twelve Calendar and chat requests, five runs each, on a computer with a 12 GB graphics card. The reply time below is the median from that test.
| Model | Right | Reply time | Download |
|---|---|---|---|
Gemma 4 (gemma4:latest) | 56 of 60 | about 3 seconds | 6.6 GB |
gpt-oss 20B (gpt-oss:20b) | 58 of 60 | about 12 seconds | 13.8 GB |
Gemma 4 26B (gemma4:26b) | 60 of 60 | about 30 seconds | 18.7 GB |
Gemma 4 small (gemma4:e2b) | 58 of 60 | very quick | 4.6 GB |
Scores alone would put Gemma 4 last, but they don’t show how a model fails. Gemma 4 asked for missing details instead of guessing them, and it answers in about three seconds, so it’s the default. gpt-oss never once broke the format of its replies, but each reply takes four times as long. Gemma 4 26B got everything right, at ten times the wait. Gemma 4 small scored well, but it put an appointment on the wrong Friday in all five runs. A user would notice that on the first day, so it’s only for computers that can’t run the others.
What PIM recommends
On first start, PIM lists these models best first and recommends the first one whose download plus 3 GB fits in your computer’s memory. With about 10 GB of memory or more, that’s Gemma 4; with less, it’s Gemma 4 small. You can pick a slower, more careful one instead, and change it later.
Every edition also uses nomic-embed-text (274 MB), a small model that finds related notes and
documents for the assistant’s memory.
One caution: a model’s tag can move to a new build. These were tested as Ollama’s registry served them on October 18 and 19, 2026, and we test again before changing any of this.
Why mistral-nemo didn’t last
We switched in August because gemma4 kept giving vague, padded answers to plain factual questions, naming where the data would come from instead of stating it. mistral-nemo answered them fully, at the same speed. That was the right fix for general questions.
Then we tested each app’s own work, and mistral-nemo fell short in most of them. In Calendar it booked a one-sentence request but asked “what title?” again on a follow-up turn. In Worklog it never offered to log the entry. In Author about half its replies came back in the wrong shape. In Mail it called an ordinary ad a scam, and its drafts signed off with “[Your Name]” when asked to sign with a name. Gemma 4 did each of those right. On the form test’s first run, on October 18, Gemma 4 got 55 of 60 and mistral-nemo 10 of 30, inventing details it hadn’t been given. On October 19, 2026, every app and general chat moved to Gemma 4: one model, one download.
If you’re upgrading from a release that used mistral-nemo, PIM preselects Gemma 4 but also offers to keep mistral-nemo, so you can update without downloading a new model. Your apps will do better on Gemma 4.