India's NandiBERT: 150M on-device encoder tops Indic benchmarks, Apache 2.0

heyshrutimishra · x · 2026-11-20

FrontiersMind open-sourced NandiBERT, an encoder trained from scratch on 11 Indian languages plus English that runs on phones (0.2s/sentence, no GPU). The 150M beats IndicBERT-v2 on 5/7 Indic tasks (67.4 vs 66.6), the 450M wins all 7 (73.2); in English both rank second in their size classes, ahead of ModernBERT. Apache 2.0, free for commercial use — enabling on-device search, support and moderation with no cloud bill.

Related event: India Open-Sources NandiBERT, a 150M Model Running on Phones in 0.2s(2 posts)→

Original post →

More from Models

Models channel →