Recipe by Mia's AI Lab @MiaAI_lab · based on Xiaomi, SGLang, NVIDIA

License: AGPL-3.0 · Xiaomi: MIT · SGLang: Apache-2.0.

TESTED BY Mia's AI Lab on 2x DGX Spark (GB10), SGLang TP=2, MXFP4 experts, EAGLE MTP · their results (benchmark table)

MiMo-V2.6-Flash on two DGX Sparks

ADVANCED: two DGX Sparks linked by a QSFP cable, Linux terminal, about 173 GB download.

FORTwo DGX Sparks. One cannot hold this model.

NEEDThe cable link, SSH from Spark 1 to Spark 2, Docker, about 161 GiB free disk on Spark 1.

WHYTop Arena WebDev score on two Sparks. 35 tok/s prose, 46 code. Text, images, video and audio in.

SPARK 1

git clone https://github.com/MiaAI-Lab/MiMo-V2.6-Flash-2x-DGX-Sparks.git
cd MiMo-V2.6-Flash-2x-DGX-Sparks
./start.sh doctor     # first run copies .env.example to .env: set the IPs there
./start.sh download
./start.sh build
./start.sh share
./start.sh serve

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 MiMo-V2.6-Flash on my two DGX Sparks: https://github.com/MiaAI-Lab/MiMo-V2.6-Flash-2x-DGX-Sparks
Follow its README exactly. Run everything on Spark 1 (the head).
1. Check first: the cable link is up, SSH from Spark 1 to Spark 2 works, Docker works, about 161 GiB free disk on Spark 1. Stop and tell me if not.
2. Clone the repo, run ./start.sh doctor, and set the IPs in .env.
3. Run ./start.sh download, build, share and serve, in that order.
4. Run ./start.sh smoke and tell me the result.
Ask me before changing network settings or deleting anything.

If it fails

A Spark runs out of memory during load.
Use the SGLang image named in the README or newer. Older engines expand the experts and take the node down.

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

What they tested

EAGLE MTP (default): 35.0 tok/s prose, 46.0 code, one request · evidence
Prefill measured at 8K, 128K and 256K prompts (764 tok/s at 256K) · evidence

Not shown in their results: The author lists concurrency scaling, long-context prefill and needle tests as not measured, though the prefill row shows 256K. We do not use it for long documents. CUDA graphs are off with the default speed-up.

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

Who made what

Xiaomi: the MiMo-V2.6-Flash model
SGLang: the serving engine (needs PR #40448 for the 4-bit experts)
NVIDIA: the Connect Multiple Sparks playbook for cables and NVIDIA Sync

From the README requirements and License section.