Comfy Worker (DGX Spark) Release Notes

This is the canonical public history for the Sogni Comfy Worker build that runs on NVIDIA DGX Spark. Each entry starts with a plain-English recap; technical details and verified Docker image information are available when needed.

DGX Spark versions and images are separate from the x86 Comfy Worker versions and images. The two families share a codebase but ship on their own schedules, so their version numbers are not expected to match and neither one waits for the other.

#Is this the right page?

This page is for a worker running on an NVIDIA DGX Spark: a Grace Blackwell GB10 machine on Linux ARM64. Everything else — every desktop and server NVIDIA card — is covered by the Comfy Worker release notes.

The installer detects a DGX Spark and selects the Spark image on its own. You do not choose it by hand.

#What this worker runs

A DGX Spark shares one pool of memory between its processor and its graphics chip, and it warms up under a long unbroken render. Left unrestricted it would accept work it cannot finish at a sensible speed, so this worker deliberately offers a small, qualified set of workflows rather than everything in the catalog:

• FastH3 — short video with sound, from a text prompt, from one picture, or between two pictures.
• MiniMax Music 3 — music generation.

Nothing else is offered on a DGX Spark, and no job for anything else will reach it. If you list an unsupported workflow in Workflows this worker serves (PREFERRED_WORKFLOW_IDS), the worker now tells you by name at startup instead of quietly dropping it.

This list will grow as more workflows are measured on this hardware. It is not a permanent judgement about what a DGX Spark can do; it is a statement about what has been tested on one.

#TL;DR for worker hosts

• The installer picks the Spark image automatically on a qualifying machine, and updates follow that image's own latest tag. You do not need to change anything after an ordinary release.
• A pinned version stays pinned. If your Compose file names an exact image such as sogni/comfy-worker-spark:1.0.209 instead of sogni/comfy-worker-spark:latest, the installer leaves it alone and you update it yourself.
• This family updates independently. An x86 Comfy Worker release does not move this worker, and a release here does not move an x86 worker.

Use dashboard.sogni.ai/fast-workers to see the version a worker is running.

#Requirements

A machine that does not meet all of these is refused rather than quietly given an x86 image it cannot run.

#Version history

#Version 1.0.209

Released: September 8, 2026 PT

Update guidance: Routine. An installer-managed DGX Spark picks this up on its next start or managed upgrade, and latest now points to it. There is no functional difference from 1.0.208 — if you are running that, nothing is wrong and there is nothing you must do.

This family now has its own version numbers. Until now the DGX Spark image borrowed its version from the x86 Comfy Worker, which meant an x86 release could renumber it and publishing a Spark image consumed a number the x86 family had not used yet. The two are now independent: this image is the first whose version comes from the Spark family's own source, and the two histories are free to diverge.

Nothing else changed. No workflow, model, memory or performance behaviour differs from 1.0.208. This release exists so the published image matches the mechanism that now owns its version.

#Version 1.0.208

Released: September 8, 2026 PT — first public DGX Spark release

Update guidance: This is the first image published for this family. An installer-managed DGX Spark picks it up on its next start or managed upgrade. There is nothing to migrate from.

A DGX Spark can run Sogni Comfy work. This is the first Comfy Worker image built for Linux ARM64 and the GB10 chip, published to its own Docker repository with its own version history.

It offers FastH3 and MiniMax Music 3. Both were measured on the hardware before being offered. Other workflows are withheld until they have been.

Memory is reported correctly for a shared-memory machine. A DGX Spark has no separate graphics memory — the graphics chip and the processor draw on one pool. The usual way of asking how much graphics memory is free reports only what is completely untouched, ignoring memory the system is holding as a cache and would release on demand. Sizing models from that number left a large model mostly unloaded on a machine with plenty of room. The worker now asks the operating system what is genuinely available, so models load in full.

Long renders are given a fair time allowance. The worker stops a render that is running pathologically slowly. That limit was calibrated on x86 machines and was short enough to cut off healthy work on this slower hardware, so it now scales with the machine.