ik_llama.cpp hits 26x speedup on Qwen 3.5

Mainline llama.cpp's struggle with hybrid SSM architectures like Qwen 3.5 highlights a growing optimization gap as linear-time models gain traction.

• –Fused GDN kernels reduce graph splits from 34 to 2, offloading recurrent computation entirely from the CPU to the GPU.
• –A 26x jump in prompt processing (from 43 to 1,122 tok/sec) makes the 27B model viable for agentic coding even with mandatory re-processing.
• –Qwen 3.5's hybrid architecture is technically superior for long context but requires specific low-level kernel support that mainline has yet to integrate.
• –Pre-built Windows binaries with CUDA 12.8 and AVX512 VNNI are available via the Thireus fork as a drop-in replacement for llama-server.