---
type: "Article"
title: "DeepSeek-V4-Flash NVFP4 with vLLM on 2× DGX Spark"
description: "DeepSeek-V4-Flash NVFP4 with vLLM on 2× DGX Spark: recipe by Weschera with speed, condition and source for every number, weights and license."
resource: "https://www.contextstudios.ai/local-ai/weschera--dsv4-flash-dspark-1m-nvfp4kv-2x-dgx-spark"
language: "en"
generated:
  by: "process:contextstudios-md/1"
  at: "2026-10-02T15:01:26.429Z"
status: "stable"
---

# DeepSeek-V4-Flash NVFP4 with vLLM on 2× DGX Spark

DeepSeek-V4-Flash NVFP4 with vLLM on 2× DGX Spark: 65.4 tok/s according to github.com (dataset as of Sep 29, 2026).

- Creator: [Weschera](https://www.contextstudios.ai/local-ai/creators/weschera.md)
- Engine: vLLM
- Quantization: NVFP4-KV-Pfad (Stage C: 584-byte padded envelope via nvfp4_ds_mla; NICHT der 416-byte true-layout Fix — dokumentiert gescheitert >411 prompt tokens)
- Model family: DeepSeek-V4-Flash
- Hardware: dgx-spark x2
- Context: 1048576

## All numbers

- 65.4 tok/s (Peak; Peak, 1 request, best value reported by source) — [source](https://github.com/Weschera/DeepSeek-v4-Flash-DSpark-1M-NVFP4-KV-2x-DGX-Spark/blob/main/README.md)
- decode_tps: 65.38 (1 request) — [source](https://github.com/Weschera/DeepSeek-v4-Flash-DSpark-1M-NVFP4-KV-2x-DGX-Spark/blob/main/README.md)

## Weights

- [fraserprice/DeepSeek-V4-Flash-DSpark](https://huggingface.co/fraserprice/DeepSeek-V4-Flash-DSpark) (main, mit)
- [deepseek-ai/DeepSeek-V4-Flash-DSpark](https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash-DSpark) (base, mit)

[https://github.com/Weschera/DeepSeek-v4-Flash-DSpark-1M-NVFP4-KV-2x-DGX-Spark](https://github.com/Weschera/DeepSeek-v4-Flash-DSpark-1M-NVFP4-KV-2x-DGX-Spark)

## Related

- [Local AI on your hardware.](https://www.contextstudios.ai/local-ai.md)