ggml-org/llama.cpp is an LLM inference project implemented in C/C++. The b12122 release adds an MMQ path that fuses the CUDA gate operation, the up-projection matrix multiplication, and the gate linear operation in NVFP4. It also covers W4A4 and W4A8, along with dense/sparse mixture-of-experts paths. The specific changes in this update are published in the b12122 release notes on GitHub.
The core of this release is that three operations in NVFP4 are fused into a single MMQ path, and that W4A4/W4A8 and dense/sparse mixture-of-experts paths are covered alongside it. The public measurements of the related change report improved per-model throughput on several devices, including DGX Spark and RTX 5090. The measured results therefore span multiple devices and models rather than being limited to a single machine. However, the provided information does not specify the individual models or the measurement conditions in detail.
That the public measurements target the related change matters for interpretation. No grounds are given to assert that the throughput improvement arose from b12122's NVFP4 MMQ fused path alone. The size of the per-model improvement is also absent from the published information. Therefore this result should not be read as meaning that the same improvement appears across all models or devices. The scope that the release notes confirm reaches only as far as the report of per-model throughput improvements on several devices for the related change, and the benefit in an individual environment is hard to judge from this information alone.