Dev trains 47M-parameter East African language model from scratch, runs offline on phones
AgencyInside407 · reddit · 2026-11-13
An indie dev shared BULaMU, a family of language models (20M/47M/110M params) trained completely from scratch for low-resource East African languages. Over recent months the corpus was expanded 10x to 40B tokens, covering Swahili, Amharic, Somali, Kinyarwanda, Lao, and English — East African languages make up 90%. A 47M base model was fine-tuned for cross-lingual story generation, and the models are small enough to run offline on phones without a GPU.
- Live demo: tinyafricanstories.streamlit.app
- Base and fine-tuned models are open on Hugging Face
Related event: BULaMU: Tiny Models Bring Offline AI to East African Languages(2 posts)→
More from Research
- DeepMind used LLMs to write 6M lines of Lean proofs, verifying Signal's crypto to machine code — pushmeet · 2026-11-13
- General Reasoning: $750K to prove out the West's first open multi-agent RL recipe for LLMs — rosstaylor90 · 2026-11-13
- Yang Songlin on how attention sink differs from gated attention: logZ and stacking — SonglinYang4 · 2026-11-13
- Gemini 4 Hydrogen probed 52,923 times reveals 239,266 brands in its parametric memory — dejanseo · 2026-11-13
- UC Berkeley's ATLAS drives humanoid robots from a single human video, no VLA or task-specific training — ChongZzZhang · 2026-11-13
- muP Lets You Tune Hyperparameters on a Small Model and Transfer to Giant Ones — RobertTLange · 2026-11-13