dgx-spark-qwen38
GitHub

dgx-spark-qwen38 / Models / Flash FP4-Local

Run Flash FP4-Local on a DGX Spark

The FP4 build of the flash model, by NVIDIA. On vLLM.

NVIDIA's FP4 build of the Flash-Next, on vLLM: the flash lane's quality in about half its disk, with twice its KV cache. On an ASUS Ascent GX10 with 128 GB: 55.4 tok/s, greedy median, at its default effort.

Measured

55.4tok/s

greedy median, at its default effort

  • 106.6 tok/s in all at 8 streams, at the lean effort
  • GSM8K 97.64% on 1,314 problems
  • About half the disk of the flash lane
Measured on
an ASUS Ascent GX10 with 128 GB, the project's reference box
How
./bench.sh at the xhigh effort
Source
docs/flash-local-lane.md

Load it

From the cockpit: open the Lanes view, pick Flash FP4-Local and press Load (17 to 19 minutes to answer the first time, about 5 and a half after). From a terminal, in the repo folder on the box:

./prepare-lane.sh flash-local   # the first time only, builds the lane
./switch-model.sh flash-local   # switches the boot, then prints the commands to run

switch-model.sh never restarts an engine itself: it prints the exact stop and start commands. The cockpit's Load runs them for you.

On a 64 GB DGX Spark

No: it needs a 128 GB DGX Spark. A 64 GB box serves Nemotron 3.5 Bolt and Qwen3.8-27B: what runs on 64 GB.

Use it from your tools

The box serves OpenAI- and Anthropic-compatible APIs on port 30001, with an API key. Claude Code, opencode, Open WebUI and the OpenAI and Anthropic SDKs point there instead of a cloud: see Claude Code with a local model and opencode with a local model.

Install it

One command, on stock DGX OS. The same command updates it later.

curl -fsSL https://raw.githubusercontent.com/hasso5703/dgx-spark-qwen38/main/get.sh | bash

Questions about Flash FP4-Local on a DGX Spark

Can a DGX Spark run Flash FP4-Local?

Yes. NVIDIA's FP4 build of the Flash-Next, on vLLM: the flash lane's quality in about half its disk, with twice its KV cache. The installer of dgx-spark-qwen38 sets it up, and the cockpit loads it with one button.

How fast is Flash FP4-Local on a DGX Spark?

55.4 tok/s, greedy median, at its default effort, measured on an ASUS Ascent GX10 with 128 GB. 106.6 tok/s in all at 8 streams, at the lean effort. GSM8K 97.64% on 1,314 problems. About half the disk of the flash lane.

Does Flash FP4-Local run on a 64 GB DGX Spark?

No: it needs a 128 GB DGX Spark.

The other lanes