Recipe by Tom Sarihan @tsarihan · based on tonyd2wild, Z.ai, RedHatAI, vLLM, NVIDIA

License: Apache-2.0 · Z.ai: MIT · RedHatAI: MIT · vLLM: Apache-2.0.

TESTED BY Tom Sarihan on 2x DGX Spark (GB10), NVFP4, vLLM TP=2, MTP-4 · their results (log or results file)

GLM-5.3-Flash on two DGX Sparks

ADVANCED: two DGX Sparks, Linux terminal, about 185 GB download per box.

FORTwo NVIDIA DGX Spark, linked by a QSFP cable.

NEEDThe two Sparks networked first. NVIDIA playbook: Connect Two Sparks.

NEEDDocker. 185 GB free disk on each Spark.

WHYA model too big for one Spark. Tested at 21 tok/s.

GET (BOTH)

git clone https://github.com/tsarihan/glm-5.3-flash-nvfp4-2x-dgx-spark-playbook.git
cd glm-5.3-flash-nvfp4-2x-dgx-spark-playbook
hf download RedHatAI/GLM-5.3-Flash-NVFP4 --local-dir /data/models/glm-5.3-flash-nvfp4

PREP (BOTH)

./scripts/prelaunch-gb10.sh

SPARK 1

NODE_RANK=0 NODE0_FABRIC=<spark-1 cable IP> ./scripts/serve-glm53-nvfp4-vllm.sh

SPARK 2

NODE_RANK=1 NODE0_FABRIC=<spark-1 cable IP> ./scripts/serve-glm53-nvfp4-vllm.sh

CHECKcurl http://<spark-1>:8890/health. Not /v1/models: that can say OK while the engine is dead.

AFTER REBOOTRun the PREP line again on both.

SAFETYIt serves an API on port 8890 with no password. Keep the Sparks on a network you own.

OR · LET A CODING AGENT DO IT

PASTEInto Claude Code, Codex or Cursor:

Set up the GLM-5.3-Flash two-DGX-Spark recipe by Tom Sarihan: https://github.com/tsarihan/glm-5.3-flash-nvfp4-2x-dgx-spark-playbook
Follow its README and scripts exactly. I have two DGX Sparks.
1. Check first and tell me: both Sparks reachable over SSH, the QSFP link up (NVIDIA Connect Two Sparks playbook), Docker working, 185 GB free on each. Stop if any is missing.
2. On both: clone the repo and download RedHatAI/GLM-5.3-Flash-NVFP4 to /data/models/glm-5.3-flash-nvfp4.
3. On both: run ./scripts/prelaunch-gb10.sh.
4. Start rank 0 on Spark 1 and rank 1 on Spark 2 with ./scripts/serve-glm53-nvfp4-vllm.sh, NODE0_FABRIC set to Spark 1’s cable IP.
5. Poll /health on port 8890 (never /v1/models) and tell me when it is up.
Ask me before changing network settings or deleting anything.

If it fails

/v1/models says OK but nothing answers.
The engine died. Check /health and the container log, then restart both ranks.
It hangs after a reboot, or long prompts freeze a box.
Run ./scripts/prelaunch-gb10.sh on both again. Its memory settings do not survive a reboot.
The engine refuses to start with images on.
Expected at full context. Leave MAX_MODEL_LEN at the script default, or set MM_IMAGES=0.
Answers cut off and look broken.
Raise max_tokens in your client. It is a reasoning model and thinks first.

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

What they tested

21.45 tok/s single stream; 39.96 / 49.30 / 61.29 tok/s at 4 / 8 / 16 users · evidence
Needle-in-haystack 5/5 at 4K, 32K and 131K tokens · evidence
SWE-bench Pro, 40-task subset: 36/40 · evidence

Not shown in their results: Uses a prebuilt container image from tonyd2wild (ghcr.io). The full 731-task SWE-bench Pro run was not finished when we checked.

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

Who made what

tonyd2wild: the original 2x Spark recipe and the container image this one starts from
Z.ai: the GLM-5.3-Flash model
RedHatAI: the NVFP4 quantized weights
vLLM: the serving engine
NVIDIA: the Connect Two Sparks playbook for the cable and network

The README Credits section names Z.ai, RedHatAI and vLLM. The serve script names tonyd2wild’s recipe and image as its starting point.