Step-3.7-Flash
StepFunReleased under Apache 2.0 with BF16, FP8, NVFP4, and GGUF weights on Hugging Face
specific method · filed under inference & serving
NVFP4 stores each value in four bits (E2M1, fifteen distinct values from -6 to 6) and gives every block of 16 values one FP8 scale, plus one 32-bit scale for the tensor, which comes to 4.5 bits per value. The finer block and the non-power-of-two scale take twice MXFP4's scale overhead and put the scale closer to each block's actual maximum.
Written for the explainer Quantization for serving: which bits to remove, and what each removal saves.
Also called NVFP4 4-bit floating-point quantization, NVFP4 checkpoint, NVFP4 (NVIDIA 4-bit floating point), NVFP4, NVFP4 quantization checkpoint, NVFP4 quantized checkpoint serving.
4 quotes from 4 sources, one card per model, each opening with its strongest.
Released under Apache 2.0 with BF16, FP8, NVFP4, and GGUF weights on Hugging Face
We recommend using the NVFP4 checkpoint nvidia/Qwen3.5-397B-A17B-NVFP4 for optimal serving efficiency on GB200 nodes.
Nemotron 3.5 Lightning ships with an NVFP4 checkpoint alongside BF16, using the same specialized NVFP4 kernels that power Nemotron 3 Ultra across NVIDIA Blackwell, NVIDIA Hopper, and NVIDIA Ampere GPUs.
NVIDIA's ModelOpt NVFP4 checkpoint
Picked by hand, not extracted: where to read more, not evidence for anything on this page.