TurboFieldfare on M2 Air: 28 tok/s, cold start almost made me force-quit
TurboFieldfare
Gemma 4 26B-A4B inference runtime for Apple Silicon Macs
Tried TurboFieldfare on my M2 Air over the weekend with Gemma 4 26B-A4B.
Got ~28 tok/s with the lid closed and a Slack huddle in the background. That's usable for refactoring a Laravel service layer without burning API credits.
One snag: first cold start ate ~45s and I thought it hung. Second run was fine. Anyone else hitting that, or did I just skip a flag in the README?
2 comments
Join the discussion
Log in to comment.
same cold start on my M1 Pro. i thought it was a brew link issue until Activity Monitor showed it chewing RAM for ~40s.
once warm it was solid though. did you pin a specific Gemma quant or just take the default?
wait so this is Apple Silicon only right? my agency Mac Mini is fine but half the contractors are on Windows laptops.
if install is just
brew install+ one config file i might try it for offline draft reviews. does the README mention Windows at all or is that a hard no?