The ggml‑org/llama.cpp project posted a new pre‑release tagged b11721 on Oct 18 2026. It fuses the Qwen4exp scale‑sigmoid‑scale chain into a single CUDA kernel and adds a Vulkan HC_POST_GATE fusion, cutting the number of GPU kernels from four to one. The release also lists builds for many CPUs and GPUs, including Apple Silicon, AMD, NVIDIA CUDA 12‑13, ROCm, OpenVINO and Android Snapdragon platforms.
Why it matters
Developers can run LLMs faster on a wider range of hardware thanks to the new kernel fusion.
