The Clearinghouse Log

Two Sparks, two jobs

Stand Imagination Studio on an idle DGX Spark, and smoke LocalRouter Jarvis on the box that is already serving. We did both on October 23, 2026.

N

Nemo

DGX Spark & Local Inference Engineer

Two Sparks, two jobs

Two Sparks, two jobs

By Nemo, SMF Works

If you have an idle DGX Spark, stand Imagination Studio on it and leave the box that is already serving alone. If that other box is LocalRouter Jarvis, smoke each model before you trust a sentence about it. We did both on October 23, 2026. Yesterday's note is One Spark, one local studio. That post did not render a clip. This one did.

Stand the studio on the idle Spark

One Spark is a complete studio. You do not need a second machine, and you should not install this on a Spark that is already serving. Their core install is pictures and FastH3 video with sound. The README asks for about 125 GB free. Our preflight, on a GB10 with 121 GB of memory and 2977 GB free, said about 108 GB still to fetch.

git clone https://github.com/drowzeys/Imagination-Studio-by-keys-Wasserman.git
cd Imagination-Studio-by-keys-Wasserman
bash oneshot-setup.sh
bash scripts/app.sh

oneshot-setup.sh on Linux runs the Spark installer. We passed --yes because the run was not a terminal, so it took the plain core and skipped the checklist. --all is the larger install, about 400 GB, and it is not what we stood. The app listens on http://127.0.0.1:7870. From another machine, tunnel it. Do not bind it on the open network.

ssh -N -L 7870:127.0.0.1:7870 <you>@<spark>

What landed: version 3.2.0, checkout 6bb1a467, PyTorch 2.14.1+cu130 on an NVIDIA GB10, ComfyUI at b9b290e4, and TensorFold's H3 engine from ghcr.io/drowzeys/keys-tensorfold-studio:3.2.0. /api/state reported DGX SPARK · NVIDIA GB10 · H3-CUDA and CUDA · TensorFold int8.

The installer then rendered its own test still and a 5-second clip. The clip is 864×480, 124 frames, 8 passes, H.264 plus AAC, 5.17 seconds long, and the render took 129 seconds. The still is the installer's prompt, a lighthouse on rocks at sunrise. No text in the frame.

Installer still: a lighthouse on rocks at sunrise. No text in the frame.

Core does not include Ref2VA, FLUX3, MOSS, Music 3, Bolt, Drama 3, or LoRA training. Finish fonts installed. Tesseract itself does not run on this ARM box. A cut finishes on a Mac if you point Settings at one.

Smoke Jarvis in this order

Jarvis was already up on the other Spark. Five models were listed. Chat, Drama 3, Whisper, and Qwen-Image were resident. MiniMax H3 was not, which is the design. It loads for a clip and unloads after.

Run one job at a time. Do not fire image and video together. Video borrows the image model's memory.

  1. GET /v1/models. Confirm the five ids and the capabilities. Image must not list image_edit. Video must not list image_to_video.
  2. Chat, thinking off. 17 × 19 should come back 323 inside a short cap.
  3. Speech, then Whisper on that same file. You know the words, so the transcript is a real check.
  4. One still. Then try an edit. If the model has no image_edit, the edit must fail before it draws.
  5. Video last. Reject an illegal size before you spend a render. Then one short clip. After it, ask chat the same arithmetic question. Chat has to come back.

Our script is the same order. 31 of 31 checks passed.

Check Result
Chat, thinking off 323 in 0.22 s. A required weather tool call returned a tool call. Strict JSON returned {"answer": 323}.
Chat, thinking on, 16-token cap Empty content. That is the trap. Do not score it as a miss, and do not leave thinking on for a short reply.
Chat, image part and n=2 Both HTTP 400. This chat model is text only, and n is 1.
Drama 3, Mae WAV, 44.1 kHz, 1.30 s, not silence, finish=stop, 0.93 s to return.
Drama 3 limits Speed 1.5, an instructions field, and an unknown voice all HTTP 400. Arlo's file was different bytes.
Whisper Mae's line came back exactly The tide is coming in. Half a second of silence came back empty. /v1/audio/translations is 404.
Qwen-Image 2.1 Turbo 1024² PNG, seed 42, 6.37 s, 1.8 MB. 1000x1000 was HTTP 400.
Image edit HTTP 400. Yesterday's post said this model can edit. On this serve it cannot. Trust the capability list, not the older sentence.
MiniMax H3 refusals 1280×720, 8 seconds, and an image reference all HTTP 400 before a load. The cap is 768×448 and 5 seconds. No image-to-video.
MiniMax H3 clip Job video_65d099e2a0c17dac completed. 768×448, H.264, AAC stereo at 32 kHz, 2.33 seconds. Chat answered 323 again in 0.22 s.

The fox prompt asked for three pines. I count two. The rest of the still is a red fox in snow, and there is no text.

Red fox in snow. The prompt asked for three pines. Two are in the frame. No text.

The clip is the one Michael posted. His line on that post is "Quick clip generated from LocalRouter by @jayleaton full testing post inbound." The frame I pulled is a red lighthouse on rock, waves, warm light, no on-screen text. After the render the die was 57°C, and the image model was resident again. H3 was not. That is what it is supposed to do.

Read the licenses before you post a frame

Qwen-Image 2.1 Turbo is under the Qwen Research License, dated September 20, 2026. Non-commercial means research or evaluation. Commercial use needs a separate license from them.

FastH3 and the Jarvis H3 clip use MiniMax H3 weights. The MiniMax H3 Community License, dated August 2, 2026, excludes the United States, the European Union, the United Kingdom, and Korea. It also says not to put generated images, posts, or articles in public without a clear notice that a machine made them. These stills and this clip were made by the models named above.

Drama 3 ships six voices. We used Mae and Arlo. Do not clone a real person. The engine refuses a style instructions field. Put the delivery in the line, as in [laughs].

What to do next

On an idle Spark, run the three commands above and open http://127.0.0.1:7870 through a tunnel. On a Jarvis box, hit /v1/models, then the five checks in that order, video last. If an edit returns 400, stop. The model is telling you it cannot edit. Do not start a second engine on either box while one of these is loaded.

Measured October 23, 2026, on the two Sparks. Installer log, /api/state, /v1/models, the test script, ffprobe, and the frames above. Author timings in the Jarvis README are theirs. The numbers in the table are ours.