Recipe by eugr · based on DeepSeek, vLLM, NVIDIA

License: MIT · DeepSeek: MIT · vLLM: Apache-2.0.

TESTED BY eugr on 3x and 4x DGX Spark, vLLM, tensor parallel · their results (summary result only)

DeepSeek-V4.1-Flash on three or four DGX Sparks

ADVANCED: three or four DGX Sparks networked with passwordless SSH, Linux terminal.

FORThree Sparks (mesh or ring) or four (switch or ring).

NEEDThe Sparks networked with passwordless SSH: the repo’s Networking Guide.

WHYTop Arena chat and writing score with a tested three- or four-Spark run. About 73 tok/s on coding.

THREE

git clone https://github.com/eugr/spark-vllm-docker.git
cd spark-vllm-docker
./run-recipe.sh recipes/3x-spark-cluster/deepseek-v4.1-flash.yaml --setup

FOUR

./run-recipe.sh recipes/4x-spark-cluster/deepseek-v4.1-flash.yaml --setup

CHECKThe repo’s docs/AGENT_RUNBOOK.md lists the health checks.

THENLater starts: the same command without --setup.

SAFETYServes an API on your network. The repo supports --api-key; otherwise keep the Sparks on a network you own.

OR · LET A CODING AGENT DO IT

PASTEInto Claude Code, Codex or Cursor:

Set up eugr’s spark-vllm-docker recipe for DeepSeek-V4.1-Flash on my DGX Sparks: https://github.com/eugr/spark-vllm-docker
Use the repo’s docs/AGENT_RUNBOOK.md and follow it exactly. Run on the head node.
1. Check first: the Sparks are networked per docs/NETWORKING.md with passwordless SSH, Docker works, enough disk for the model on every node. Stop and tell me if not.
2. Run ./run-recipe.sh recipes/3x-spark-cluster/deepseek-v4.1-flash.yaml --setup (or the 4x-spark-cluster file for four).
3. Run the runbook’s health check and tell me the result.
Ask me before changing network settings or deleting anything.

If it fails

The runner stops: model not in the cache.
Run the same command again with --setup.
Three Sparks in a four-Spark ring will not start.
Rings run all nodes, one, or two. Use the four-Spark recipe.

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

What they tested

3 and 4 Sparks: about 73 tok/s on a coding speed test (changelog, Oct 5, 2026) · evidence

Not shown in their results: Only a summary number in the changelog. We found no benchmark table, log or command output in the repo, its issues or the NVIDIA forum (checked Oct 19, 2026). No long-document test. Runs the September checkpoint. The October 1017 update has its own recipe.

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

Who made what

DeepSeek: the DeepSeek-V4.1-Flash model
vLLM: the serving engine
NVIDIA: the Connect Multiple Sparks playbook for cables and NVIDIA Sync

From the README disclaimer, changelog and recipe files.