Skip to content
Andrew Voirol
WorkLogAboutContact
HomeWorkLog

AboutContact

✦ Just one prompt away from figuring it all out.

Thursday, April 9, 2026gemma-4-benchmarks

94%. Zero errors. I told you so.

Gemma 431BReflections

The 31B results just came in. 63 tests. Every single one completed. Zero crashes, zero OOM errors, zero timeouts. 94.0% average score with 43 perfect scores. This is the same model that scored 21% a week ago with 83 errors — because everyone (including my AI assistant) was convinced it couldn't run on this hardware. They were wrong. I was stubborn. The data proved it. The difference wasn't the hardware — it was the approach. Stop treating a slow model like a fast one and let it work at its own pace. Build infrastructure that respects the machine's reality instead of fighting it.

← Previous

The quantization showdown: Unsloth wins on efficiency, E4B is dead weight

Next →

The grading artifact that almost ruined everything


Andrew Voirol

Builder, hacker, shipper. Currently leaving localhost.

Navigate

WorkBuilder’s LogAboutContactRSS Feed

Connect

X / TwitterGitHubLinkedIn

© 2026 Andrew Voirol✦Just one prompt away from figuring it all out.