llama.cpp lands native NVFP4 on Blackwell

This is a meaningful infrastructure win for local AI on Nvidia’s newest hardware: the format support is in place, and the bench numbers suggest it is already useful for real workloads, not just a compatibility checkbox.

• –Native NVFP4 matters because Blackwell’s 4-bit path is part of the hardware story, not an afterthought; llama.cpp is now tracking that story closely.
• –The benchmark profile is heavily prefill-friendly: 5.5K+ tok/s at short contexts, then a gradual drop as depth increases, which is what you’d expect from a memory-bandwidth and attention-pressure story rather than a pure compute bottleneck.
• –Decode around 70 tok/s on a 27B model is strong for a local setup, especially with the whole model on a single GPU and no CPU offload in the test.
• –This is a good signal for Blackwell owners, but it is still one data point from one model and one build; different architectures, contexts, and batch shapes can change the picture.
• –The release note framing suggests this is the first real integration step, so expect follow-up fixes and tuning as more NVFP4 models and kernels land.