Hacker News Viewer

TinyLoRA – Learning to Reason in 13 Parameters

by sorenjan on 3/27/2026, 12:11:12 PM

https://arxiv.org/abs/2602.04118

Comments

by: a-t-c-g

The quality of custom models trained with proper reasoning datasets[0] even with small parameters (3-7B is sweet spot) is incredible now<p>[0]: cartesien.io or Salesforce&#x27;s WebscaleRL

4/1/2026, 2:01:15 AM


by: measurablefunc

With four parameters I can fit an elephant, and with five I can make him wiggle his trunk so there is still room for improvement.

4/1/2026, 1:08:22 AM


by: ValveFan6969

[dead]

4/1/2026, 1:56:39 AM