June 2026 focused on hardening the model conversion workflow for ggml-org/llama.cpp, with an emphasis on robustness and reliability of the Mistral-Medium-3.5-128B conversion path. The changes reduce downstream issues for users migrating models to gguf and improve overall stability of the conversion tooling.

February 2026 performance month focused on CUDA dequantization optimizations across iq2xxs/iq2xs/iq3xxs for two key repos (ggml-org/llama.cpp and ggml-org/ggml). Delivered low-level data-path improvements that reduce latency and improve throughput on relevant hardware by optimizing how dequantization data is loaded and how signs are computed. Key techniques: - Load all 8 int8 values for a grid position in a single load - Compute signs via popcnt instead of fetching from a signs table - Broadcast signs to drop per-element shifts/masks, simplifying the path Impact: - Reduced register usage in the critical mul_mat_vec_q path (152 -> 149), enabling better occupancy and potential throughput gains (nsight-confirmed). - Consistent improvements across both llama.cpp and ggml, aligning performance characteristics across repos and hardware targets. This work is captured in dedicated commits linked to the dequantization optimization effort (llama.cpp and ggml) to (#19624) style references and mirrors across repositories for consistency.