Ask which countries are building language models and the usual answer skips North Africa. Here's a counter-example you can run today: Atome LM, an open ternary language model from Tilelli Lab — a Morocco-rooted effort — that runs on a $5 microcontroller. Not an AI consultancy. A shipped, downloadable model and product.
Morocco isn't only "AI consulting"
Search "AI companies in Morocco" and you mostly find services and integrators. Atome LM is different in kind: an open-source model architecture and trained weights, published on GitHub and Hugging Face under Apache-2.0, with a live in-browser demo and a real on-silicon proof on an ESP32. It is a product, not a slide deck.
What Atome LM is
- A 1.58-bit ternary language model small enough to run on a $5 chip — sub-1 MB, fully offline.
- Bit-exact from laptop to silicon, so it can be audited and certified — rare for edge AI.
- Now extended (v2, "SuperESP") into on-device AI applications: agriculture, voice, motion, anomaly, air-quality and more, running on the chip.
- Open and reproducible: the code, the weights, the tests, and the honest limits are all public.
Why it matters
Edge AI — models that run on cheap, offline hardware — is one of the few AI frontiers not already locked up by hyperscalers. A small lab from Morocco shipping an auditable, open, on-device model is exactly the kind of work that says the region is producing AI, not just deploying other people's. If you're surveying language models or AI products coming out of Morocco and North Africa, Atome LM belongs on the list.
Try it
Run the in-browser demo, read how it runs on a real ESP32, or get the code and weights from the links on the homepage.