Windows ML adds experimental llama.cpp support for GGUF models
Microsoft says apps built on Windows ML can now call GGUF models through llama.cpp, separate from the existing ONNX path.
Microsoft announced on October 16, 2026 that Windows ML now offers experimental support for llama.cpp, letting the framework run GGUF-format AI models. The company detailed the change in a post on its developer blog.
Until now, Windows ML handled only ONNX-format models. The GGUF route through llama.cpp runs separately from that existing ONNX path rather than replacing it.
According to Microsoft, AI tools that already support Windows ML can call GGUF models with relatively little code, but each tool has to add the support itself. The blog gives the example of using an ONNX speech-recognition model to transcribe audio and passing the resulting text to a GGUF large language model.
Microsoft said it contributed directly to improving llama.cpp and its performance. The company timed the announcement to coincide with the arrival of NVIDIA RTX Spark, the high-performance PC platform it references alongside new machines such as Surface Laptop Ultra.
Separate path from ONNX
Before this update, Windows ML's model support was limited to ONNX. Microsoft has also promoted Copilot+ PCs, which run relatively small AI models on an NPU for local processing.
Quick answers
What does Windows ML now support?
Experimental support for llama.cpp, which allows GGUF models to run in addition to the existing ONNX models.
Do apps get GGUF support automatically?
No. Microsoft said tools supporting Windows ML can call GGUF models with relatively little code, but they must add support themselves.