Skip to main content
Back

NVIDIA Releases DeepSeek-V4.1-Flash Checkpoint Optimized for Blackwell GPUs

NVIDIA released a 4-bit NVFP4 build of DeepSeek-V4.1-Flash that cuts memory to ~492 GiB while matching source accuracy on Blackwell GB300.

Model, Full Access
  • NVIDIA released DeepSeek-V4.1-Flash-NVFP4, a 4-bit quantized build of the 552B-parameter multimodal MoE model.
  • Uses W4A4 NVFP4 on routed experts with group size 16; attention and vision stay at higher precision.
  • Checkpoint size is roughly 492 GiB across 48 safetensors shards, up from 476 GiB in MXFP4 source.
  • Accuracy matches source within noise on GPQA Diamond, SciCode, IFBench, MMMU-Pro, and Terminal-Bench 2.1.
  • Requires Blackwell GPUs; validated on four GB300s with vLLM and SGLang at 1M context.
  • Built with NVIDIA Model Optimizer v0.47.0rc0; MIT license, commercial use permitted.

NVIDIA has published a pre-quantized DeepSeek-V4.1-Flash checkpoint using its NVFP4 4-bit floating-point format. The release targets Blackwell-class GB300 systems and includes tested configurations for vLLM and SGLang, allowing teams with compatible hardware to avoid running their own calibration and quantization pipeline.

DeepSeek-V4.1-Flash is a multimodal Mixture-of-Experts model with a causal encoder-decoder architecture, Compressed Sparse Attention 2, and a maximum context length of 1,048,576 tokens. Its 552-billion-parameter backbone activates 8 billion parameters during prefill and 16 billion during decoding. Another 196 billion parameters provide Engram conditional memory. Hugging Face metadata counts about 763 billion parameter entries across the model’s mixed tensor types.

NVFP4’s scaling scheme

NVFP4 stores each quantized value in E2M1 format, using one sign bit, two exponent bits, and one mantissa bit. It applies two levels of scaling: an FP8 E4M3 scale for every block of 16 values and an FP32 scale for the entire tensor. The additional scaling information helps preserve values that would otherwise be distorted when compressed to four bits.

Compared with MXFP4, NVFP4 uses smaller scaling blocks of 16 elements instead of 32 and a more precise FP8 E4M3 scale instead of E8M0. Those choices can reduce quantization error, although they also increase metadata and restrict native acceleration to NVIDIA hardware that supports FP4 Tensor Cores.