Oriel — AI & technology worth noticing, every day.

nvidia/DeepSeek-V4.1-Flash-1017-NVFP4

AI model · https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-NVFP4

First seen by Oriel on .

AI summary

NVIDIA’s DeepSeek-V4.1-Flash-1017 NVFP4 is a pre-quantized, multimodal MoE checkpoint for Blackwell deployment, reducing storage while maintaining benchmark results close to the MXFP4 source model.

  • The model has 552B total backbone parameters, 8B activated during prefill, 16B during decode, and 196B Engram conditional memory.
  • NVFP4 converts routed-expert weights and activations to W4A4 with group size 16, while Engram lookup tables use lossy MXFP4 conversion.
  • The checkpoint occupies approximately 400 GiB across 48 safetensors shards, compared with approximately 475 GiB for the source.
  • Deployment requires a vLLM build with MXFP4 Engram support, four NVIDIA GB300 GPUs in the reported setup, and Linux.
  • NVFP4 plus MXFP4 Engram scored 92.93 GPQA Diamond, 80.19 AA-LCR, 57.32 SciCode, 76.47 IFBench, 78.73 MMMU-Pro, and 81.11 Terminal-Bench 2.1.

#deepseek #mixture-of-experts #model quantization #vllm

Generated by gpt-5.6-sol. Check the original for details.

Metric trends

Last 30 days · Up to 120 recorded samples per source and metric, with reposts kept separate · First and latest retained

Likes

Total likes reported by the source.

  • Hugging Face Models: 184, latest observation ; +79 over observed period since

Downloads (30d)

Rolling 30-day downloads. A change in this value is not new downloads during the observed period.

  • Hugging Face Models: 7,821, latest observation ; +6,348 over observed period since

Aliases

SourceExternal IDURL
Hugging Face Modelsnvidia/DeepSeek-V4.1-Flash-1017-NVFP4https://huggingface.co/nvidia/DeepSeek-V4.1-Flash-1017-NVFP4 Read discussion