The ggml‑org/llama.cpp project released version b12327 on 12 Nov 2026. The update adds a SpacemiT Q4_1 32×256 kernel that uses the same high‑performance (HP) kernel as Q4_0. It also finishes the M1 (Apple Silicon) code path, allowing the new quantization to run on macOS arm64. The release lists dozens of platforms, from Linux CPUs and GPUs to Windows CUDA 13 and Android Snapdragon.
Why it matters
Developers can now run smaller, faster LLaMA models on Macs and other devices with less memory usage.
