FTerViT: A Fully Ternary Vision Transformer That Runs on a $10 Microcontroller
Prior ternary ViTs leave the patch embedding, LayerNorm, and classifier head in full precision — on tiny models those leftovers dominate the memory budget. FTerViT ternarizes everything to {-1, 0, +1}: 82.43% ImageNet top-1 at 6 MB, on a $10 ESP32-S3.
— Research
