When people picture machine learning they imagine racks of GPUs. But some of the most useful models in the world run on chips that cost less than a cup of coffee and sip microamps of power.
The memory budget is the design
On a microcontroller with 256 KB of RAM, every kilobyte is a decision. TinyML forces you to think about the model and the system together:
- Quantize aggressively — 8-bit integer weights cut memory by 4x with minimal accuracy loss.
- Stream your features — never hold more of the input in memory than you need for the next inference.
- Fuse operations so intermediate tensors never fully materialize.
Why the edge matters
Running inference on-device means no round trip to the cloud: lower latency, no data leaving the sensor, and it keeps working when the network does not. For agriculture, health, and industrial monitoring across Africa, that offline-first reality is not a nice-to-have — it is the whole point.
A realistic pipeline
Train in the cloud, convert to a quantized format, and deploy the resulting flatbuffer straight to flash. The same model that trained on a workstation now lives on a chip you can hold between two fingers.