llama.cpp b9549 adds Gemma4 MTP for faster local inference

llama.cpp, the widely-used open-source C/C++ inference engine for large language models, released version b9549, adding support for Google's Gemma4 with Multi-Token Prediction (MTP). Gemma4 is Google's recent lightweight family of models designed for efficient on-device AI. MTP is an advanced inference technique that allows the model to predict multiple tokens in a single forward pass, significantly boosting generation throughput compared to standard autoregressive decoding. This release integrates that capability, allowing users to run Gemma4 with up to 2x faster output on local hardware without cloud dependencies.

The release provides pre-built binaries for all major platforms: macOS (Apple Silicon with optional KleidiAI acceleration, and Intel x64), Linux (x86/arm64 CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (x64/arm64 CPU, CUDA 12/13, Vulkan, HIP), Android (arm64 CPU), and iOS via XCFramework. Users can also build from source. The inclusion of MTP aligns with the broader trend of making state-of-the-art inference optimizations accessible to the developer community, further cementing llama.cpp as a key tool for running AI models locally.