April 2026 (2026-04) focused on delivering targeted kernel performance improvements in the ggml-org/llama.cpp project, with emphasis on the Q4_MMQ kernels (q4_0 and q4_1). The main feature delivered was a performance optimization that replaces ds_read_b32 with ds_read_b128 to reduce LDS bandwidth and enable faster loads, accompanied by vectorized loading updates and loop-level refinements. This work included explicit loop restructuring and fixes to the loading loop, and a typo correction in the q4_1 kernel. In addition to feature work, code quality improvements were applied, including cleanup in mmq.cuh and removal of trailing whitespace. The changes were validated on multiple GPU platforms (MI50 and RX6800XT) and are documented in the merge commit 66c4f9ded01b29d9120255be1ed8d5835bcbb51d, with co-authors contributing to cross-platform validation. Overall, the month delivered tangible performance gains for critical inference kernels, improved reliability of the loading path, and reinforced code quality and collaboration practices.