Documented local result
Flash-Next IQ3_XXS
llama.cpp · IQ3_XXS
41/42quality
16.92tok/s decode
UNRAID LAB / 128 GB DDR4 / RTX 3060
The fast model isn’t the reliable one.
Measured local inference, deterministic grading, and the failures that matter. No vendor numbers. No LLM judges.
Current verdictQwen for speed. Flash-Next for strict work.
RUN INDEX / 003
Latest evidence
Initial results are note-derived and clearly marked. Missing raw artifacts limit reproduction, not disclosure.
Documented local result
Qwen3.6 NVFP4
FreeToken · NVFP4
7/9quality
61.5tok/s decode
Documented local result
Flash-Next Q3_K_XL
llama.cpp · Q3_K_XL
9/9quality
14.5tok/s decode
FINDING 01
Speculation doesn’t beat a memory bottleneck.
MTP accepted roughly 0.8 of drafted tokens but produced no wall-clock improvement on CPU-offloaded Flash-Next.
Read the finding →0%measured speedup