Early Inkling 2 Mini LoRA tests: same results with under a third of the energy

ziqiao_ma · x · 2026-11-21

A developer training Inkling 2 Mini on a DGX Spark reports that LoRA fine-tuning only six layers matches results from all 24 layers while using less than a third of the energy.

ziqiaoma adds that the model is a good candidate for an OOD autoresearch benchmark variant for PostTrainBench or InferenceBench, noting big optimization headroom in looped transformer inference and post-trainability.

Original post →

More from Infra

Infra channel →