Qwen3.5 4B Q4_0

Run (UTC)	geist	llama.cpp	pp512 tok/s	vs llama	tg64 tok/s	vs llama	Files
2026-10-09 13:42	53aacaacb	d81235049	256.9	-18 %	26.0	+13 %	report · raw

Every run of this cell, newest first. pp512: prefill of 512 tokens; tg64: decode of 64 tokens at depth 512; median of 12 samples. The raw file holds every sample and the full host, engine and model provenance. Back to the matrix.