The llama.cpp project has released version b12122, which adds a CUDA optimization that fuses NVFP4 gate/up matrix multiplications with a GLU operation in MMQ, its matrix-multiplication path. NVFP4 is a compact number format used in some AI computations; combining steps can reduce separate work, though the release notes give no performance measurements.
The release also includes downloadable builds for macOS, iOS and Ubuntu on x64, Arm64 and s390x.
Why it matters
People running supported NVFP4 workloads on CUDA may get the fused computation path, but the release notes do not quantify any speed or memory gains.
