After finishing some refactors and improvements, I've updated my #anvilope testbed machine with gemma 4 E2B. It's been running for about two days, and going just off feels the results seem quite nice. The categorizing seems pretty similar to gmail's built-in inbox categories, which was the entire point of this project. Needless to say I'm feeling pretty good!
The previous models I used on testbed were mistral 7B and qwen3 4B. Mistral was ok but slower, and qwen3 was... kind of excitable and inconsistent. All in all I'm pretty impressed with this gemma 4 model so far.
after I let this config run for a while I should also try one of the small nemotron models