llama.cpp fixes Qwen4-exp tensor shape bug

How it developed

1. October 24, 2026llama.cpp ReleasesWhere it started

   llama.cpp fixes Qwen4-exp tensor shape bug

2. October 26, 2026llama.cpp Releases

   llama.cpp fixes 4D tensor quantization for MoE models

   llama.cpp fixes a critical bug in llama-quant that silently dropped data when processing 4D tensors in modern Mixture of Experts architectur

3. October 28, 2026llama.cpp Releases

   llama.cpp decouples sparse attention from mask shape

   llama.cpp decouples k-pool sparse selection from dense kq_mask shape to support Qwen4exp and GLM5-next architectures