Spark Stack

One command turns one or two NVIDIA DGX Sparks (or GB10 boxes such as the ASUS Ascent GX10) into a local AI lab: a web app to download, start and stop models, two agent harnesses already pointed at them, and one API address that always reaches whatever model is running.

curl -fsSL https://raw.githubusercontent.com/sybil-solutions/spark-stack/main/install.sh | bash

Run it on the Spark you sit at, as your normal desktop user; it asks for your password once. With a QSFP cable to a second Spark (same username and password on both, both on the same Tailscale network) it finds that Spark and sets it up too. Running the same command again repairs an install. ~/spark-stack/uninstall.sh removes it (weights are kept unless --weights).

Models

The fastest public recipe for each model and Spark count, all MiaAI-Lab TensorFold recipes, pinned by commit in stack/registry.py. Speeds are the recipes' own single-user measurements.

Model	1 Spark	2 Sparks	Download	Port
Qwen3.8-Flash-Next (180B-A6B, vision)	Single-DGX-Spark-TensorFold: ~50 tok/s	Dual-DGX-Sparks-TensorFold: 67 tok/s	96 / 124 GiB (NVIDIA NVFP4)	8889
Qwen3.8-27B (dense, vision, DFlash2 drafts)	27B-DGX-Spark-TensorFold: 58 prose, 137 code	not needed	19 GiB	8891
GLM-5.3-Flash (320B-A18B, vision)	not offered (2-bit only, 10-25 tok/s)	EXL3-2x-DGX-Sparks-TensorFold: 60 prose, 115 structured	166 GiB	8888
DeepSeek-V4.1-Flash-1017	does not fit	1017-2x-DGX-Sparks: 75 prose, 110 code	385 GiB	8890

Install downloads Qwen3.8-Flash-Next (two-Spark build, or the one-Spark build on a single Spark) and starts it as soon as the download is verified. Everything else downloads when someone presses Download or Start. One model runs at a time. A start takes about 70 seconds once the weights are on disk.

What the installer does

1. Finds the second Spark: on the QSFP cable, else the one tailnet machine with a live CX7 link that answers Tailscale SSH; WORKER=user@host overrides, SINGLE=1 skips it.
2. On both Sparks (scripts/node-setup.sh): turns sleep off, adds the user to docker, addresses both logical twins of the cabled CX7 port (192.168.100.10/11, 192.168.101.10/11), installs rsync and NFS, and shares DeepSeek's weights-only cache (~/.cache/spark-stack-hf) read-only to the worker's CX7 addresses.
3. Installs the hf CLI (both Sparks), Node.js and the DeepSeek harness (/opt/spark-stack-node, service spark-stack-dsh, web UI on 127.0.0.1:3080), and the OpenCode desktop app.
4. Clones each recipe at its pinned commit into ~/spark-stack/vendor/.
5. Builds and starts the spark-stack container (restarts with the machine), bound to this Spark's Tailscale address only: http://<tailscale-ip>:8099.
6. Writes the models into the DeepSeek harness (~/.dsh/profiles/web/cordis.patch.yml, managed block) and OpenCode (~/.config/opencode/opencode.json, provider spark), and pins Spark Stack, DeepSeek Harness and OpenCode to the dock.

The app

Spark Stack in the dock opens it in its own window (desktop/webapp.py, GTK WebKit; DeepSeek Harness opens the same way, already signed in). From another tailnet machine it is http://<tailscale-ip>:8099: paste the key from ~/spark-stack-state/token once. Each model card has a 1 Spark / 2 Sparks choice, Download with a progress bar, Start/Stop, and the live log. Disk lists the model folders on both Sparks with a Remove button.

After a reboot or power cut, the model that was running starts again by itself (for a two-Spark model, once the second Spark is back). Stop opts out: a stopped model stays stopped.

One API for every app

http://<tailscale-ip>:8099/v1 is OpenAI-compatible and needs no key. Model id spark goes to whatever is running; each model's own id (Qwen3.8-Flash-Next, Qwen3.8-27B, GLM-5.3-Flash-EXL3, DeepSeek-V4.1-Flash-1017) goes to that model and gets a clear 503 when it is not running.

method	path	purpose
GET	/api/state	machines, models, download progress, what is running
POST	/api/models/{key}/prepare?sparks=1|2[&start=true]	download (start=true: then start it)
POST	/api/models/{key}/load?sparks=1|2	start (downloads what is missing first)
POST	/api/models/{key}/unload	stop
GET	/api/models/{key}/logs	the job log
GET	/api/disk	model folders on each Spark
POST	/api/disk/remove	{"path": ...} remove a model folder from both Sparks
GET	/api/stats	live tokens a second, power, memory

Keys: qwen-next, qwen-27b, glm, deepseek.

Recovery

• Something looks stuck: Stop the model in the app, then Start it again.
• The app does not open: sudo docker restart spark-stack.
• Anything else, or after an OS update: run the install command again.

Test

On the head, with weights downloaded: tests/e2e.sh qwen-next 2 starts the model, chats through /v1 by id and by spark, checks a stopped model returns 503, and stops it.

Credits

Recipes, engines and images: MiaAI-Lab (TensorFold). Weights: NVIDIA, Vontra, z-lab, Mia-AiLab, incoai, DeepSeek. Harnesses: DeepSeek harness, OpenCode.