MORENA: 1.5B African LLM trained from scratch on $40K of GPUs beats Google and Meta models
letandrewcook · x · 2026-09-18
- Researchers trained MORENA, the largest African language model ever built from scratch (1.5B parameters) — not a fine-tune, but trained ground-up for Africa.
- Trained on 12 African languages, it beats models from Google, Meta, Alibaba and Hugging Face on African language modelling and translation while being up to 8x smaller.
- Total GPU cost: about $40,000, with support from UNDP, AIHub4SD and CINECA.
- Weights are open and free for commercial use, aiming to give researchers a base model that doesn't treat African languages as an afterthought.
More from Models
- How TypeSafe Jev turns LLM token probabilities into snap judgments — and why frontier labs could clone it — JnBrymn · 2026-09-18
- Cryptographer mocks OpenAI: sandboxes can't even run a recent Linux kernel — matthew_d_green · 2026-09-18
- GPT-6 weekly limits spawn 'usage maxing': users race the reset timer — rschu · 2026-09-18
- Researcher: real-time driving-game models were doable in 2016 — industry has LLM tunnel vision — gandamu_ml · 2026-09-18
- Altman says Sora's decline comes down to one thing: 'not enough compute' — rohanjamin · 2026-09-18
- Matt Wolfe Weekly: Musk, Altman, Hassabis Back Dario's AI Pacing Essay, Zuckerberg Pushes Back — Matt Wolfe · 2026-09-18