llama.cpp b9654 adds post-decode callback with AI-assisted coding

ggml-org has tagged llama.cpp b9654, a maintenance release that introduces a post-decode callback within the multi-token matching decoder (mtmd). Announced via the commit e3cab40, the new callback allows users to inject custom processing after each token is decoded, enabling use cases like real-time moderation, streaming adjustments, or analytics logging. The release stands out for being assisted by the pi:llama.cpp/Qwen3.6-27B model itself, an interesting example of using an AI to help develop the infrastructure that runs such models.

The release builds are available for nearly every major platform: macOS (Apple Silicon and Intel), Linux (x64, ARM64, s390x with multiple compute backends like Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (CPU, CUDA 12/13, Vulkan, HIP), and Android ARM64. iOS and XCFramework are also provided. This breadth ensures that developers running local LLMs on everything from a Raspberry Pi to a multi-GPU workstation can benefit from the improved hooking capabilities. The callback is a small but impactful step for those building custom pipelines around llama.cpp, making it easier to extend inference behavior without forking the entire library.