Physix Frontier · News Briefing Card (enbrief · Sep 24, 2026)

Vambo AI Releases 1.5B-Parameter MORENA Model Covering 12 African Languages

KEY FACTS

Vambo AI has released the MORENA model, with 1.5 billion parameters, covering 12 African languages plus English, French, and code.
MORENA achieved the best score of 1.408 bpb among 26 tested models, beating rivals from Google, Meta, Alibaba, and others.
The model was built from scratch, not trained further on existing models such as Llama, and its vocabulary is designed specifically for African languages.
Development consumed over 22,000 A100 GPU hours, with computing resources worth about $40,000, and received support from institutions including the UNDP.
The team also released smaller 0.5B and 0.2B versions; the 0.2B version generates 104 tokens per second and can run offline on a laptop.

KEY DATA

1.5BModel parameters

1.408Best bpb score

251.7BPretraining tokens

22000GPU hours

PHYSIX OBSERVATION

African languages have long been marginalized by mainstream models, and MORENA proves that a small model plus a dedicated vocabulary can overtake the giants. It pushes training cost down to $40,000 and also offers an offline version, a pragmatic path for regions short on compute. This reminds the industry: fairness in language coverage does not necessarily have to be solved by piling on parameters.

Source: enbrief original report ↗