esp32-ai
Running a 28.9M parameter LLM on an $8 microcontroller
...This memory layout allows the device to retrieve only the embedding rows required for each token. The implementation reaches roughly 9.5 tokens per second and can display generated words on a connected screen. Trained on TinyStories, the model produces short, simple stories rather than answering questions, following instructions, or providing factual knowledge. The repository includes firmware, wiring and flashing instructions, training code, quantization experiments, ablations, and measured results.