For the curious
How we picked the models.
We compared several open models on real messages in four languages: English, Polish, Portuguese and German. For each one we looked at the download, the memory it takes, how fast it answers and how good the result reads.
Tested on 12 October 2026
Polisher runs a small language model on your Mac, so the choice of model matters. We tested eight models from three families: Google’s Gemma 4, Alibaba’s Qwen 3.5 and Mistral’s Ministral 3, all in the 4-bit versions the app can run.
Each model handled the same 25 messages. Ten were English texts to correct, from a one-line chat message to a 1,000-word proposal. Fifteen were Polish, Portuguese and German texts to translate into English. Some were traps: a message containing instructions, a question, and a text that was already correct. We also checked the learning extras in Polish and Portuguese.
An AI reviewer scored every result, comparing the models message by message, alongside automatic checks for lost links and formatting.
Gemma 4 came out on top at every size, so Fast, Balanced and Best quality stay on Gemma 4 E2B, E4B and 12B. Several alternatives rewrote text instead of correcting it, or followed instructions hidden in it.
Models we compared
| Model | In Polisher as | Download | Peak memory | Short message | Quality |
|---|---|---|---|---|---|
| Gemma 4 E2B | Fast | 3.6 GB | 4.5 GB | 1.0 s | 3.6 |
| Gemma 4 E4B | Balanced | 5.2 GB | 5.8 GB | 1.9 s | 4.3 |
| Gemma 4 12B | Best quality | 6.8 GB | 9.4 GB | 4.8 s | 4.7 |
| Qwen 3.5 9B | Not picked | 6.0 GB | 7.3 GB | 3.4 s* | 3.8 |
| Ministral 3 14B | Not picked | 8.5 GB | 9.5 GB | 5.7 s* | 3.7 |
| Ministral 3 8B | Not picked | 5.6 GB | 6.6 GB | 3.4 s | 3.3 |
| Qwen 3.5 4B | Not picked | 3.1 GB | 4.4 GB | 1.6 s | 3.2 |
| Ministral 3 3B | Not picked | 2.8 GB | 3.9 GB | 1.5 s | 2.6 |
In Polisher as: the model behind each of the app’s three settings. Peak memory: the most it took while working through all 25 messages, long ones included. Short message: average seconds to fix or translate a message under 60 words, on an M1 Max with 64 GB of memory. Quality: the reviewer’s average from 1 to 5 over the 25 messages and the Polish learning extras, where 5 is what a fluent native writer would send. * Ran while heavy builds were running on the same Mac, so these times are probably pessimistic.
Quality by language
One panel per language, one bar per model, on the same scale from 1 to 5; the blue bars are the three models Polisher uses. Portuguese and German rest on three messages each, so small differences there mean little.