Recipe by Mia's AI Lab @MiaAI_lab · based on TensorFold, Z.ai, turboderp, IncoAI, Jay Leaton, NVIDIA

License: Apache-2.0 · TensorFold: Apache-2.0 · Z.ai: MIT · IncoAI: CC BY-NC-ND 4.0, non-commercial.

TESTED BY Mia's AI Lab on 2x and 3x DGX Spark (GB10), EXL3 4-bit, DFlash2, 1M context · their results (benchmark table)

GLM-5.3-Flash on two or three DGX Sparks (TensorFold)

ADVANCED: two or three DGX Sparks linked by QSFP cables, Linux terminal, about 176 GB download.

FORTwo DGX Sparks. Three also work (experimental).

NEEDThe Sparks cabled and networked first, with SSH from Spark 1 to the others.

NEEDAbout 110 GiB free memory on each Spark at start.

WHYTop Arena chat, coding, writing and images score on two Sparks. 60.4 tok/s on two, 77.6 on three.

NOTEThe default drafter (DFlash2) is non-commercial only. DRAFTER=mtp avoids it, one request at a time.

SPARK 1

git clone https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold.git
cd GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold
cp scripts/local.sh.example scripts/local.sh
# set WORKER=user@<spark-2 address> in scripts/local.sh
./start.sh

THREESet WORKER2 too, then run ./start-tp3.sh instead of ./start.sh.

READYIt prints “GLM-5.3-Flash-EXL3 is now LIVE! on port 8888”.

TESTcurl -s http://<spark-1>:8888/v1/models

STOP./stop.sh

SAFETYThe API on port 8888 has no password in the README. Keep the Sparks 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 GLM-5.3-Flash on my DGX Sparks: https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks-TensorFold
Follow its README exactly. Run everything on Spark 1 (the head).
1. Check first: the QSFP link is up and pings, SSH from Spark 1 to the worker works, Docker works, about 110 GiB free memory on each. Stop and tell me if not.
2. Clone the repo, copy scripts/local.sh.example to scripts/local.sh and set WORKER (and FABRIC_PEER if the README says so).
3. Run ./start.sh (or ./start-tp3.sh with WORKER2 for three Sparks).
4. Wait for “is now LIVE! on port 8888”, test /v1/models and tell me.
Ask me before changing network settings or deleting anything.

If it fails

“only N GiB memory available”.
Other GPU work runs on that Spark. Stop it (docker ps) and restart.
“no RoCE device” or “no route”.
The route to the worker skips the cable. Set FABRIC_PEER to the worker’s cable address.
Answers cut off.
It thinks first. Give replies more max_tokens.

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

What they tested

Two Sparks: 60.4 tok/s prose, 114.7 structured, one request; 108.8 total at 4 · evidence
Two Sparks: a 981,841-token needle found; GSM8K 98.8%, HumanEval 95.7% · evidence
Three Sparks (experimental): 77.6 tok/s prose, one request; 195K needle right · evidence

Not shown in their results: Measured with the author’s own tool (sparkDash). Three Sparks are marked experimental by the author.

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
Z.ai: the GLM-5.3-Flash model
turboderp: the EXL3 format and converter (exllamav3)
IncoAI: the DFlash2 drafter (speed-up)
Jay Leaton: tool calling, prefetch and expert-load code (glm53-tensorfold-spark)
NVIDIA: the Connect Multiple Sparks playbook for cables and NVIDIA Sync

From the README Credits and License sections.