---
title: "Discussion on \"Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s\" | Hashnode"
description: "Discussion on \"Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s\". I benchmarked Qwen3.6-35B-A3B on DGX Spark GB10 using llama.cpp (45–50 tok/s) vs vLLM + NVFP4 + DFlash (88–104 tok/s). Full setup guide, docker-compose, troubleshooting, and real-world results."
url: "https://hashnode.com/posts/qwen3-6-on-dgx-spark-vllm-nvfp4-dflash-vs-llama-cpp-2x-faster-at-88-104-tok-s/69fb2aa950ecad45331b9b3a"
---

[Feed](https://hashnode.com/feed) /

Discussion

# [Qwen3.6 on DGX Spark: vLLM + NVFP4 + DFlash vs llama.cpp — 2x Faster at 88–104 tok/s](https://devsnack.hashnode.dev/qwen3-6-on-dgx-spark-vllm-nvfp4-dflash-vs-llama-cpp-2x-faster-at-88-104-tok-s?utm_source=hashnode&utm_medium=feed)

I benchmarked Qwen3.6-35B-A3B on DGX Spark GB10 using llama.cpp (45–50 tok/s) vs vLLM + NVFP4 + DFlash (88–104 tok/s). Full setup guide, docker-compose, troubleshooting, and real-world results.

May 6 · 10 min read · devsnack.hashnode.dev  
Read the article

-   [#nvidia](https://hashnode.com/tag/nvidia)
-   [#dgx-spark](https://hashnode.com/tag/dgx-spark)
-   [#dgxspark](https://hashnode.com/tag/dgxspark)
-   [#qwen](https://hashnode.com/tag/qwen)
-   [#qwen3-6](https://hashnode.com/tag/qwen3-6)
-   [#qwen-3-6](https://hashnode.com/tag/qwen-3-6)
-   [#vllm](https://hashnode.com/tag/vllm)
-   [#llamacpp](https://hashnode.com/tag/llamacpp)
-   [#llama-cpp](https://hashnode.com/tag/llama-cpp)
-   [#nvfp4](https://hashnode.com/tag/nvfp4)
-   [#speculative-decoding](https://hashnode.com/tag/speculative-decoding)

## Written by

[강문규](https://hashnode.com/@devsnack)

@devsnack · Seoul, South Korea

  
[View profile](https://hashnode.com/@devsnack)

Nvidia dgx sparks gb10, Asus gx10, llm, local ai

Software developer from South Korea.

## Responses

Markdown supported · ⌘↵ to post

No responses yet.