Recipe by Mia's AI Lab @MiaAI_lab · based on 0xSero, knapcio, DeepSeek, SGLang, NVIDIA

License: AGPL-3.0-or-later · 0xSero: MIT · DeepSeek: MIT · SGLang: Apache-2.0.

TESTED BY Mia's AI Lab (four-Spark line by knapcio) on 3x and 4x DGX Spark (GB10), SGLang, MXFP4 experts, DSpark · their results (benchmark table)

DeepSeek-V4.1-Flash-1017 on three or four DGX Sparks

ADVANCED: three or four DGX Sparks, Linux terminal, 476 GB checkpoint.

FORThree Sparks in a cable triangle, or four on a switch.

NEEDThe cable links, SSH between the Sparks, Docker, and NVMe space for the Engram tables on each Spark.

WHYThe newer DeepSeek update. Top Arena WebDev score of all models here. 87.7 tok/s on four Sparks, 37.9 on three.

CONFLICTThree Sparks: the README gives 51.0 tok/s in its summary and 37.9 in its comparison table. It also says 208K prompts verified and about 32K usable. We show the lower figures.

THREE

git clone https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks.git
cd DeepSeek-v4.1-Flash-DGX-Sparks
cp -n .env.example .env    # set IPs and users
./start.sh doctor && ./start.sh download && ./start.sh build
./start.sh share && ./start.sh pack && ./start.sh serve

FOUR

cp -n .env.tp4.example .env.tp4    # set IPs
./start-tp4.sh doctor && ./start-tp4.sh build && ./start-tp4.sh share
./start-tp4.sh pack && ./start-tp4.sh serve

FIRST BOOT12 to 13 minutes.

STATUS./start.sh status · ./start.sh smoke · ./start.sh stop

SAFETYSet API_KEY in .env so that the server asks for a password. Without it, 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 DeepSeek-V4.1-Flash-1017 on my DGX Sparks: https://github.com/MiaAI-Lab/DeepSeek-v4.1-Flash-DGX-Sparks
Follow its README exactly. Use start.sh for three Sparks or start-tp4.sh (docs/tp4.md) for four.
1. Check first: cable links, SSH between the Sparks, Docker, disk and NVMe space. Run ./start.sh doctor and stop if it fails.
2. Set the IPs in .env (or .env.tp4), then run download, build, share, pack and serve in that order.
3. Run ./start.sh smoke and tell me the result.
Never set OFFLOAD_MODE=ram. Ask me before changing network settings or deleting anything.

If it fails

The head runs out of memory and all ranks stall.
Never use OFFLOAD_MODE=ram. Keep other programs off the head Spark.
Replies run forever.
Always send max_tokens. The server also stops repeating loops.

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

What they tested

Four Sparks: 87.7 tok/s prose, 124.8 code, one request; 342.7 total at 16 · evidence
Four Sparks: a 1,011,084-token needle passed · evidence
Three Sparks: 37.9 tok/s prose, one request, about 32K usable context (comparison table). The summary table says 51.0 tok/s and 208K prompts verified. · evidence

Not shown in their results: Three Sparks: the README lists 51.0 tok/s in its summary but 37.9 in its comparison table, and its context notes conflict (208K prompts verified vs about 32K usable). We do not use the three-Spark run for long documents. Measured with the author’s own tool (sparkDash).

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

Who made what

0xSero: the MIT-licensed recipe this one builds on
knapcio: the four-Spark production line (PR #36)
DeepSeek: the DeepSeek-V4.1-Flash-1017 model
SGLang: the serving engine
NVIDIA: the Connect Multiple Sparks playbook for cables and NVIDIA Sync

From the README License section and the four-Spark credits.