ggml-org has released llama.cpp b12029, a new version of its software for running language models. The release notes list one change: a fix to the order of InternVL’s CLS token, a special token used when processing multimodal input. Downloads are listed for macOS, iOS and several Linux configurations, including CPU and Vulkan builds. The release notes do not explain which users or tasks are affected by the fix.
Why it matters
People running InternVL through llama.cpp may benefit from the token-order fix, but the release notes do not specify its practical effect or scope.
