The ggml‑org team released llama.cpp version b11890 on Oct 26 2026. The update cleans up DeepSeek V4 code, moves some model parameters into a central config, and drops unused model classes. It also expands hardware support, adding CUDA 12 and 13 libraries, Vulkan, ROCm and OpenVINO backends for Linux, Windows and macOS. The changes let developers run larger models faster on more GPUs and NPU‑enabled devices.
Why it matters
Developers can now run DeepSeek V4 models on a wider range of GPUs, reducing the need for cloud‑based inference.
