llama.cpp b9354 adds MiniCPM5 tokenizer for local inference

The open-source llama.cpp project, known for efficient local inference of large language models, released version b9354 on May 27. The headline feature is tokenizer support for MiniCPM5, a small yet capable language model from OpenBMB. This integration allows developers to convert and run MiniCPM5 models locally using llama.cpp's optimized C++ backend. The tokenizer uses a BPE (Byte-Pair Encoding) approach with a hardcoded regex, consistent with other pre-tokenizers in the project. Co-authored by Zhang Tao from ModelBest, the update ensures smooth tokenization for MiniCPM5's vocabulary.

Beyond the tokenizer addition, the release provides pre-built binaries for a wide range of platforms: macOS (Apple Silicon with optional KleidiAI, Intel x64), iOS (XCFramework), Linux (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32), Android ARM64, and Windows (CPU, ARM64, CUDA 12/13, Vulkan, HIP). This broad support means developers can deploy MiniCPM5 on everything from cloud servers to edge devices. The release also includes UI assets for the project's built-in web interface. Users can now leverage MiniCPM5's efficient architecture for tasks like text generation, summarization, and code completion directly on their own hardware, without relying on cloud APIs.