Comfy Worker (DGX Spark) Release Notes

The Sogni Comfy Worker for NVIDIA DGX Spark is new, and support is growing as each workflow is measured on this hardware. This page is the canonical public history for its releases.

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 serves a deliberately qualified set of workflows. As of 1.0.211, that includes:

• FastH3 short video with sound, from text, one picture, or two endpoint pictures.
• MiniMax Music 3 music generation.
• Krea 2, Dark Beast Krea 2, and their identity-edit workflows.
• Z-Image Turbo, Dark Beast Z-Image Turbo v9, and One Obsession v22 image generation.
• Chroma v.46 Flash, Chroma v.48 Detail Calibrated, and Chroma1-HD image generation.
• Qwen Image Edit 2511 Lightning image editing.
• SAM 3 interactive segmentation and BiRefNet background removal.
• Pixal3D image-to-3D reconstruction and GLB export.
• Qwen3-TTS custom voices, voice cloning, and voice design.

Workflow admission remains one workflow at a time. The same list controls the model download plan, so a Spark fetches the files for work it can serve and does not fetch models for unqualified workflows.

Optional model packs work differently on Spark after qualification: if an opt-in workflow is on the Spark admission list, its files download automatically. Hosts upgrading from 1.0.210 should allow roughly 130 GB for the one-time download behind the workflows added in 1.0.211. Files live in the worker's persistent host model cache, so an ordinary container restart, image update, or worker upgrade does not download them again. A newly provisioned host with no existing cache must download its complete targeted set.

#What to expect from the hardware

A DGX Spark shares one pool of memory between its processor and graphics chip. It holds its clocks well across the short jobs this worker serves and falls back under long, unbroken load. Plan sustained throughput on the throttled rate rather than the rated one, and give the machine ordinary airflow and room. The worker's temperature guard pauses new work at 87 °C and resumes at 79 °C, as it does on every other GPU.

#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.
• A pinned version stays pinned. If your Compose file names an exact image such as sogni/comfy-worker-spark:1.0.211 instead of sogni/comfy-worker-spark:latest, the installer leaves it alone.
• This family updates independently. An x86 Comfy Worker release does not move this worker, and a release here does not move an x86 worker.
• Updating works the same way as on any other worker. Use Upgrade worker in the Dashboard or run ./worker-restart.sh, and prefer an idle moment over interrupting active work.

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

#Requirements

An ARM64 machine that does not meet all of these is refused rather than quietly given an x86 image it cannot run.

#Version history

#Version 1.0.211

Released: September 9, 2026 PT

Update guidance: Recommended. An installer-managed DGX Spark picks this up on its next start or managed upgrade. Allow time and disk space for the one-time model download described above.

Thirteen more workflows are now offered. The release adds Z-Image Turbo, Dark Beast Z-Image Turbo v9, One Obsession v22, three Chroma variants, Qwen Image Edit 2511 Lightning, SAM 3 segmentation, BiRefNet background removal, Pixal3D, and three Qwen3-TTS speech workflows.

Qualified opt-in packs now download automatically on Spark. SAM 3, BiRefNet, Pixal3D, and Qwen3-TTS normally require an operator to opt in on other Comfy workers. Once their workflows passed Spark qualification and joined this hardware's admission list, Spark began including their files automatically. Downloads go into the persistent model cache rather than the container image.

Speech support includes the complete runtime. The ARM64 image now includes Qwen3-TTS and its CPU ONNX tokenizer dependency. Speech generation still runs on the GPU; this does not broaden support to unrelated ONNX workflows.

#Version 1.0.210

Released: September 9, 2026 PT

Update guidance: Superseded by 1.0.211. Upgrade to the current release.

Four Krea 2 image workflows were added. A DGX Spark began serving Krea 2 Turbo, Dark Beast Krea 2, and both Krea 2 Identity Edit variants alongside FastH3 and MiniMax Music 3. All four were measured on the hardware first and completed in 13 to 50 seconds of render at 832×1216, using about 19 GB of the machine's shared memory pool.

This release also caught the Spark image up with the x86 worker codebase, including the restored MiniMax H3 LightX2V Turbo LoRA and the background-removal, Pixal3D, and Qwen3-TTS code that was not yet offered on Spark.

#Version 1.0.209

Released: September 8, 2026 PT

Update guidance: Superseded. No workflow or runtime behavior changed from 1.0.208.

The Spark family began using independent version numbers. Spark and x86 releases now advance separately. The recommended version is published on its own comfy-spark endpoint, and an exact Compose pin is never overwritten.

#Version 1.0.208

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

Update guidance: Superseded. This was the first image published for the Spark family.

A DGX Spark could begin running Sogni Comfy work. The first Linux ARM64/GB10 image offered the three FastH3 workflows and MiniMax Music 3 after hardware qualification.

Shared memory was accounted for correctly. The worker began asking the operating system what memory was genuinely available instead of treating reclaimable host cache as unavailable graphics memory.

Long renders received a Spark-specific time allowance. The worker's fail-fast render limits were scaled for this machine without changing any other GPU class.