About this site
About this publication
LENA Project is an open research publication documenting the design, engineering, and results of a new African language AI architecture — written by the person building it.
What this publication is
LENA Project is the public record of building LENA — the Linear Efficient Neural Architecture — an open-source language model designed from the ground up for Oromo, Amharic, and other African languages that existing AI systems systematically underserve.
This is not a product blog or a press release. It is a documentation of real research: the design decisions, the engineering challenges, the bugs, the results, and the honest assessment of what worked and what did not. Every claim in this publication is sourced. Every result is from a completed experiment. When something did not go as planned, that is documented too.
The publication runs as a twenty-part series. It begins with the problem — why AI costs African language speakers four to seven times more per query than English speakers, and why the infrastructure to build better AI is priced beyond reach of most African institutions. It moves through the mathematics of how language models work, the engineering story of building LENA on a single consumer GPU, and the full results of the proof-of-concept training run. It ends with where the project goes next and how you can be part of it.
About the author
I am Awel Dico — Canadian with Ethiopian origin, researcher, architect and the person who designed and built LENA. My background is in both physics and software engineering: my master's thesis at the University of Bergen in 1994 developed a real-time imaging algorithm for oilfield separators using electrical capacitance tomography, and my doctoral research at Memorial University studied the molecular interactions of pulmonary surfactant proteins using Nuclear Maganetic Resonance - Quantum physics. The throughline in both — and in LENA — is the same principle: that a carefully constrained algorithm, built around the specific structure of the problem, can do more with less than a general-purpose approach given unlimited resources.
LENA is the application of that principle to the problem I care about most: making capable AI accessible to the communities that need it, in the languages they actually speak.
What LENA has demonstrated so far
The LENA-138M proof-of-concept model has completed training. In a controlled comparison against a parameter-matched dense Transformer — identical tokeniser, data, schedule, and optimiser, differing only in architecture — LENA achieved lower held-out perplexity on all three target languages: Oromo −5.1%, Amharic −6.2%, English −2.4%. The advantage concentrates on the two low-resource languages, which is exactly the regime the architecture was designed for.
LENA's tokeniser achieves a fertility of 1.61 tokens per word for Oromo and 1.62 for Amharic — comparable to what frontier tokenisers achieve for English, and three to four times better than commercial tokenisers applied to those languages. The model is designed to run inference on a 12GB consumer GPU. All weights, code, and training pipelines will be released openly.
Who this publication is for
The series is written for readers at three levels simultaneously. The problem posts (Posts 1–4) are written for anyone curious about technology and equity — no technical background required. The mathematics posts (Posts 5–10) are written for anyone with higher level secondary-school or first year university algebra and genuine curiosity — equations appear, but always with the intuition first. The engineering posts (Posts 11–18) are written for developers and researchers who want to understand how the model was actually built, including what went wrong.
If you are an African researcher, developer, or student working on language technology — this publication is specifically for you. The full training pipeline, model weights, and tokeniser will be open so that anyone can reproduce, build on, and extend this work using hardware that is realistically accessible in African institutions.
Subscribe
The series publishes weekly. Subscribing is free and gets each new post delivered directly to your inbox. No advertising, no sponsored content. The only thing in your inbox will be the next post in the series.
If you find this work valuable and want to support it — whether through funding, data contribution, compute access, or collaboration — the best first step is to subscribe and reach out directly. The contact link is at the bottom of this page.