Thirteen parameters reach 91% on a reasoning benchmark

Claim

Meta researchers pushed compression further with TinyLoRA, training a model to 91% accuracy on GSM8K with just thirteen parameters in bf16, twenty-six bytes in total, while Google introduced a video model at less than half the cost of its previous cheapest.

Reported metric: 13 parameters / 26 bytes

Reporting date: 2026-04-01 (not necessarily the event date)

Sources

Attribution and reporting linked by this record; links do not imply independent verification.

• Welcome to April 1, 2026 (theinnermostloop.substack.com)