---
title: "NVIDIA Shrinks DeepSeek V4 Flash to Run on a Single Blackwell Node"
subtitle: "NVIDIA released an official NVFP4 4-bit quantization of DeepSeek V4 Flash 0731, cutting the 284B MoE down to Blackwell-friendly size while preserving DSpark speculative decoding."
url: "https://alphasignal.ai/news/nvidia-shrinks-deepseek-v4-flash-to-run-on-a-single-blackwell-node"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "NVIDIA AI"
company_url: "https://alphasignal.ai/company/nvidiaai"
published: "2026-08-19T17:09:47.000Z"
updated: "2026-09-05T07:44:41.646Z"
type: "Model"
topics:
  - "Llms"
  - "Gpus"
subtopics:
  - "Mixture Of Experts"
  - "Quantization"
upvotes: 56
source_url: "https://huggingface.co/nvidia/DeepSeek-V4-Flash-0731-NVFP4"
comments: 0
truncated: true
---

# NVIDIA Shrinks DeepSeek V4 Flash to Run on a Single Blackwell Node

> NVIDIA released an official NVFP4 4-bit quantization of DeepSeek V4 Flash 0731, cutting the 284B MoE down to Blackwell-friendly size while preserving DSpark speculative decoding.

NVIDIA has published an [official NVFP4 quantization](https://huggingface.co/nvidia/DeepSeek-V4-Flash-0731-NVFP4) of DeepSeek V4 Flash 0731, the July refresh of DeepSeek's efficiency-tier Mixture-of-Experts model. The checkpoint is built with NVIDIA's Model Optimizer toolkit and targets Blackwell-class GPUs, where FP4 tensor cores execute 4-bit math natively rather than upcasting to FP8 or BF16.

DeepSeek V4 Flash 0731 has 284B total parameters with 13B activated per token and a 1M-token context window. Getting that into VRAM on a reasonable node is the whole problem this release addresses.

## What got quantized and what didn't

Only the routed MoE experts are quantized to NVFP4\. Attention layers, router gates, shared experts, embeddings, and DeepSeek's DSpark speculative-decoding heads all remain in their original precision. The routed experts are where almost all parameter mass lives in a sparse MoE, so quantizing them slashes on-disk and VRAM footprint substantially while leaving the numerically sensitive paths untouched.

## Why NVFP4 instead of MXFP4 or INT4

NVFP4 and MXFP4 both use symmetric per-group quantization with a fixed granularity, unlike INT4\. MXFP4 uses a group size of 32 with an FP8 (E8M0) scaling factor per group; NVFP4 tightens that to a group size of 16 with an FP8 (E4M3) scaling factor, then adds a global FP32 scale per tensor to extend the representable range. Both scaling steps are hardware-accelerated on Blackwell.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [nvidia/DeepSeek-V4-Flash-0731-NVFP4](https://huggingface.co/nvidia/DeepSeek-V4-Flash-0731-NVFP4)
- [DeepSeek-V4-Flash-0731 model card](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-0731)
- [DeepSeek-V4-Flash-0731 on Baseten](https://www.baseten.co/library/deepseek-v4-flash-0731/)
- [NVIDIA NIM: deepseek-v4-flash-0731](https://docs.api.nvidia.com/nim/reference/deepseek-ai-deepseek-v4-flash-0731)
- [Accelerating LLMs with NVFP4 quantization](https://developers.redhat.com/articles/2026/02/04/accelerating-large-language-models-nvfp4-quantization)