Quantized AI models can now be executed within the `transformers` library using GGML kernels to achieve speeds identical to llama.cpp. This update enables the direct execution of GGUF checkpoints without the performance overhead of unquantizing the files during loading. Developer Marc Sun and the GGML organization provided the kernels that power the new integration.

The transition to GGML kernels through a dedicated `kernels` library allows the Transformers ecosystem to support millions of existing GGUF downloads while maintaining the performance of the original checkpoints. Previously, Hugging Face supported GGUF loading by unquantizing the models first, which hindered the overall efficiency of the quantized files during runtime.