From-scratch 42M-parameter model handles 7 African languages, runs offline on phones
AgencyInside407 · reddit · 2026-11-13
A developer shared BULaMU, a family of from-scratch models (20M/47M/110M) for the low-resource language Luganda, small enough to run offline on a phone without GPU or internet. The corpus grew 10x to 40B tokens by adding Swahili, Amharic, Somali, Kinyarwanda, Lao and English, with East African languages making up 90%. The 47M model was fine-tuned for code-switching story generation across languages; a Streamlit demo is live and base plus fine-tuned weights are on Hugging Face (42M params, 1,024-token context, CPU inference).
Related event: Dev Trains Tiny Models for East African Languages from Scratch(3 posts)→
More from Research
- 80 Ants Beat a Team of 26 Humans at the Piano Movers Puzzle, PNAS Study Finds — aakashgupta · 2026-11-13
- Team tries Malik's 5-year-old vs robot dexterity challenge; the gap is clear — micoolcho · 2026-11-13
- Jitendra Malik's Challenge: 5-Year-Olds vs the MolmoAct2 Robot at SPIN Workshop — DJiafei · 2026-11-13
- AG2's Extra K562 Tracks Are Imputed, Not New, Clarifies Adam Gayoso — anshulkundaje · 2026-11-13
- Alignment failures share systematic errors; we need decorrelated evidence — sethlazar · 2026-11-13
- Six PIs at ELLIS Institute Tübingen Named ELLIS Fellows and Scholars — orvieto_antonio · 2026-11-13