Reverse-engineered NPU engine format runs GGUFs 1.5× faster

A developer reverse-engineered the Axera AX8850 NPU's engine format, storing int8 weights as two nibble planes, to run GGUFs without model conversion. It achieves 1.5× speedup over the vendor's runtime, running Qwen3-0.6B at 13.5–14.5 t/s on the M5Stack LLM-8850 card.