ik_llama.cpp posts huge Qwen3.5 CPU gains

This looks less like a lucky benchmark and more like proof that the local inference ecosystem is splintering into specialized forks for specific model families and hardware targets.

• –The posted numbers are hard to ignore: about 281.6 t/s vs 56.5 t/s on prompt processing and 22.4 t/s vs 12.9 t/s on token generation for the same Qwen3.5 4B quant
• –ik_llama.cpp's README explicitly emphasizes better CPU performance, custom quants, and model-specific optimizations, so the gain is consistent with the project's design goals rather than a random anomaly
• –Comments from contributors and power users point to chunked delta-net work and repeated CPU-side optimization passes as likely reasons Qwen3 and Qwen3.5 perform especially well here
• –This is still an anecdotal community benchmark, not a controlled bake-off, and at least one commenter reported weaker gains or regressions in hybrid CPU+GPU setups
• –If these results hold broadly, mainline llama.cpp risks becoming the compatibility baseline while forks like ik_llama.cpp become the speed path for serious local CPU inference