Building a CPU LLM engine in C99 - stuck at 1.90 tok/s on DeepSeek MoE while llama.cpp does 13.79. Potential root cause identified. Implementation is not.

In three linesDeveloper building LLM inference engine in pure C99. On DeepSeek-V2-Lite with i5-11300H: 1.90 tok/s vs 13.79 for llama.cpp (7.3x gap). Root cause identified: Q4K dequantization to F32 at load (4 bytes/weight) vs direct Q4K reads (0.5 bytes) in llama.cpp. IPC: 0.80 vs 2.36. Required fix: fused Q4K matvec kernel.

Summary generated by Claude — human-verified