Recipe by Mia's AI Lab @MiaAI_lab · based on TensorFold, Jürgen Schmied, NVIDIA, Qwen

License: MIT · TensorFold: Apache-2.0 · NVIDIA: NVIDIA Open Model Agreement · Qwen: Qwen Community License 1.0.

TESTED BY Mia's AI Lab on 1x DGX Spark (GB10), NVIDIA FP4 checkpoint, MTP drafts, vision on · their results (benchmark table)

Qwen3.8 Flash Next on one DGX Spark (TensorFold)

ONE SPARK: Linux terminal, Docker, about 160 GB free disk.

FOROne NVIDIA DGX Spark, nothing else big running on it.

NEEDDocker. About 160 GB free disk (96 GB model plus the engine image).

WHYTop Arena WebDev score that runs on one Spark. 46 to 55 tok/s, images and video in.

RUN

git clone https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold.git
cd Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold
./start.sh

READYIt prints “Qwen3.8-Flash-Next is now LIVE! on port 8888”.

TESTcurl -s http://<spark-address>:8888/v1/models

STOP./stop.sh

SAFETYThe API on port 8888 has no password in the README. Keep the Spark on a network you own.

OR · LET A CODING AGENT DO IT

PASTEInto Claude Code, Codex or Cursor:

Set up Mia’s AI Lab recipe for Qwen3.8 Flash Next on one DGX Spark: https://github.com/MiaAI-Lab/Qwen3.8-Flash-Next-Single-DGX-Spark-TensorFold
Follow its README exactly.
1. Check first: Docker works, about 160 GB free disk, about 108 GiB of memory available. Stop and tell me if not.
2. Clone the repo and run ./start.sh. The first run builds the engine and downloads the model.
3. Wait for “is now LIVE! on port 8888”, then test /v1/models and tell me.
Ask me before deleting anything or stopping other programs.

If it fails

It refuses to start: not enough memory.
Stop other GPU work on the Spark. The default needs about 108 GiB available at start.
TTFT looks slow from another computer.
The link adds its round trip. Measure on the Spark itself.

From the original README, shortened. Read it for full details.

What they tested

Prose 46 to 55 tok/s for one request (4 runs), code 67.9 tok/s; 132.7 tok/s total at 5 requests · evidence
Prefill measured up to a 131,110-token prompt (2,197 tok/s) · evidence

Not shown in their results: Measured with the author’s own tool (sparkDash). No needle-in-a-haystack test.

RunCards does not test recipes. We only list recipes whose authors publish their own test results. Checked 2026-10-19.

Who made what

TensorFold: the serving engine, by Ash Hart and contributors
Jürgen Schmied: the Zig CUDA serving path (TensorFold PR #443)
NVIDIA: the FP4 QAD checkpoint
Qwen: the Qwen3.8 Flash Next model

From the README Credits and License sections.