Atome LM is built by Tilelli Lab, a small AI lab in Marrakech. It is the embedded face of a larger research programme: the three-pathway architecture and the C99 engine are public, and the more elaborate variants — extended pathways, retrieval and memory, multi-bank weight schemes, an internal regression-prevention gate — are not.
Why a lab in Marrakech is working on this
Marrakech has been a meeting point of Amazigh, African, Arab and Mediterranean influences for a very long time, and that mix still shapes how the lab thinks and builds. There is also a less romantic reason. The assumption that intelligence has to live in a datacentre is an assumption made by people with cheap datacentres. Most of the world's devices are not in that position, and neither is most of the world.
How the work is run
Three habits, and they show up all over this site:
- Publish the losses. The limitations page is not an appendix. The architecture loses above a million parameters, per-window latency is a loss against CMSIS-NN, and the accuracy comparison with TFLite Micro is a tie. All three are on the site with numbers.
- A claim names the file that produced it. Not "roughly", not "up to". A figure on this site either cites the artifact it was read from or it does not appear.
- A retracted claim gets a paragraph, not a silent edit. Two latency figures and one accuracy win have been withdrawn. Each withdrawal is written down where the original claim was made.
What we do commercially
Production integration: silicon bring-up, the Atome Secure Boot Pack, per-platform hardening, domain fine-tuning and the evidence pack for regulated devices. The services page has the shape of it, and the honest version of the sales pitch is that the model is usually the easy part.
Talk to us
The useful message is a specific one: the part number, the sensor, the sample rate, the flash and SRAM budget, and what the device has to decide. Contact.