The Qwen Flash Next model now halves its indexer score memory, reducing VRAM consumption as part of pull request #29825 to the ggml-org/llama.cpp repository. This optimization improves efficiency for running the model locally.