Africa Creates Compact LLM Morena, Outperforming Google and Meta✴ Models

African company Vambo AI has introduced the Morena language model with 1.5 billion parameters. It supports 12 African languages, English, French, and can write computer code. In tasks related to African languages, this model, according to the developer, outperforms systems from Google, Meta✴, and Alibaba, while remaining eight times more compact in size than analogues.

Background and Architectural Features

At a key testing stage, Vambo AI Morena showed a result of 1.408 bpb (bits per byte), which was the best among 26 tested models. The closest to it, 1.423 bpb, was shown by a model five times larger — the lower the metric, the better. The model supports the following local languages: Nigerian Pidgin, Igbo, Yoruba, Hausa, Swahili, Shona, Zulu, Xhosa, Kinyarwanda, Tswana, Afrikaans, and Southern Ndebele. Most projects working with these languages use fine-tuned open Meta✴ Llama models, but they retain a vocabulary originally designed for English and programming code.

Technological Breakthrough in Tokenization

Vambo AI took a different approach: before training, they optimized the tokenizer, the composition of training data, and the list of languages, having previously compared vocabulary sizes in terms of cost and efficiency. When encoding African texts, the Morena dictionary uses 1.39 times fewer tokens than Google Gemma 3 and 1.53 times fewer than Llama 3.2, compared on identical fragments. Encoding costs for African text amount to 0.249 tokens per byte versus 0.234 for English — a 6% difference that developers cannot yet explain. The closest competitor is the Africa-adapted version of Meta✴'s model called Lugha-Llama-8B — it showed 1.423 bpb, supporting only 8 instead of 12 languages. For comparison, the 12-billion-parameter Gemma model consumes about 11 times more computational resources than Morena.

Instruct Version Efficiency and Development Costs

Morena optimized for conversational interaction (the instruct version) demonstrates a metric of 1.441 bpb, trailing Lugha-Llama-8B overall, but outperforming it on five shared languages while using only 20% of its parameters. When translating texts from English into five African languages, the instruct version shows 45.8 on the chrF++ metric, which is statistically comparable to specialized translation systems — their result is about 1.4 points higher, while models comparable in size to Morena usually show from 9 to 14 points. During pre-training, Morena processed 251.7 billion tokens, plus another 63 billion during intermediate training.

Model Lineup and Availability

The development required over 22,000 hours of Nvidia A100 accelerator operation — computational resources cost about $40,000. In addition to the main 1.5-billion-parameter model, 0.5-billion and 0.2-billion versions are available. The 0.5-billion parameter variant still outperformed all tested models from other developers across 11 out of 12 languages. The most compact 0.2-billion parameter version is designed for re-evaluating options in speech recognition systems, keyboard applications, and text normalization tasks. The latter costs 58 times less than Lugha-Llama-8B per byte, and when running on a single Nvidia A100 accelerator, it generates 104 tokens per second. A version capable of running on a CPU and locally on a laptop is also available. Vambo AI Morena supports chatbot applications, translation features, and integration with other AI tools.