Glossary
NVFP4
Nvidia's 4-bit floating-point weight format, used to train and serve very large models on Blackwell-class GPUs with roughly a quarter of the memory of 16-bit weights.
Inkling 2 Small was trained natively in NVFP4 and shipped that way: 594 billion parameters in about 394 GB. The catch is hardware: NVFP4 kernels target Blackwell (B200, B300, GB300) and need conversion or special kernels on older or non-Nvidia accelerators, which is why 'open weights' and 'runnable on your cluster' are not the same thing.
Related terms
Closed modelComputeConfidential computingExpert parallelismGPU
Get the brief in your inbox
One email every morning: the stories that mattered in AI, each with a one-line "why it matters" for your role. Free, no ads, unsubscribe any time.
Published daily by AI Briefing. We never share your address.