Needle3 on the ESP32-S3 just hit 4.78 tok/s. We started at 1.22 tok/s on the unmodified 8-layer baseline. That's now: * 3.92× the original decode speed * +292% performance * Prefill: 1.25 → 5.22…

For 1:1 with JEV needle doesn't compare out of the box. I learned they want you to finetune it for specific use case. We know that small models can do great work for finetuned. But I had a different…