DMX¶

Development line: project:dmx · thread dmx-format-release
Last event: 2026-04-09 · 1 dated since 2026-04-09 · Researched: 2026-09-05 · confidence: medium

What it is¶

DMX is a Python package for teams distributing safetensors models: lossless .dmx archives, base-relative .dmxd deltas, round-trip verification, and embedded lineage metadata. Limit: every PyPI release listed through 1.5.3 is yanked, so it is not a safe unpinned production dependency.

Development line¶

• 2026-04-09 — DMX was publicly presented as an AI model-weight compression format. On 2026-04-09, DMX was publicly presented through a ComfyUI community post linking a GitHub repository and a Hugging Face profile. The linked materials identify the project as concerning compression of AI model weights. The available evidence does not establish its technical specifications, release maturity, or supported integrations.

What changed¶

2026-04-09 — DMX launched as a CLI-oriented compressor for model storage and transfer; the release-day package history records versions 0.4.0 and 0.5.0, later yanked. The original workflow required decompression back to safetensors before ComfyUI use; the launch cited 9.1 GB to 1.8 GB, 7.2 GB to 1.5 GB, and +0.16% perplexity on Llama 3 8B. 2026-04-10 — PyPI version 0.6.0 was released and later yanked. 2026-04-11 — PyPI version 0.7.0 was released and later yanked. 2026-04-19 — PyPI versions 1.0.0 and 1.1.0 were released and later yanked; the later package documentation describes lossless archives, deltas, provenance, and a GPU runtime. 2026-04-20 to 2026-04-25 — versions 1.2.0 through 1.5.3 were released and all later yanked. A Qwen2.5-3B-Instruct delta demonstrates the newer distribution path: 2.88 GB int16 delta versus a 13.59 GB full model, but it requires the exact Qwen2.5-3B base and a merged single-file export.

How to use this¶

As of 2026-04-09, practitioners evaluating model-weight storage or transfer can treat DMX as a newly announced compression-format project, but should verify implementation details and compatibility in its repository before adopting it.

1. Do not install an unpinned current release: first select and audit a specific wheel because the listed releases are yanked. — https://pypi.org/project/dmx-compress/
2. For a lossless archive, run dmx compress model.safetensors model.dmx, then prove it with dmx verify model.safetensors model.dmx. — https://pypi.org/project/dmx-compress/
3. Restore a portable checkpoint with dmx decompress model.dmx restored.safetensors; use the restored safetensors in tools that do not natively load DMX. — https://pypi.org/project/dmx-compress/
4. For the published Qwen delta, obtain the exact Qwen/Qwen2.5-3B base, apply dmx delta-reconstruct, then load the reconstructed safetensors in the target framework. — https://huggingface.co/Senat1/dmx-qwen2.5-3b-instruct-delta

Best practices¶

• Keep the original checkpoint and run a round-trip verification before deleting or distributing a DMX artifact. — https://pypi.org/project/dmx-compress/
• Treat a delta as bound to its stated base checkpoint; reconstruction is expected to fail with a different base. — https://huggingface.co/Senat1/dmx-qwen2.5-3b-instruct-delta
• For the April 9 workflow, budget DMX as disk and download compression, not ComfyUI VRAM reduction; native ComfyUI loading was only a roadmap item then. — https://www.reddit.com/r/comfyui/comments/1saeezm/i_built_a_compression_format_for_ai_model_weights/

Superseded by this¶

• 2026-04-09 — Guidance that DMX was only an external storage compressor is outdated relative to later package claims of compressed-residency runtime support, but those claims remain insufficiently reproducible for production adoption.
• 2026-04-09 — Guidance to expect a native ComfyUI node is obsolete as a current instruction: use restored safetensors unless a separately verified integration is available.

Still unknown¶

• The requested response schema has no event_findings or new_events fields; the event-specific additions and later dated events are therefore represented in what_changed.
• The original GitHub repository URL returned 404 when checked, so repository code, changelog, and current maintainership could not be independently audited.
• The PyPI README reports runtime benchmarks and support claims, but its public technical and benchmark documentation is marked as forthcoming; treat those measurements as maintainer claims rather than independently reproduced results.

Sources¶

Agent brief¶

• Subject: project:dmx, thread dmx-format-release, 1 dated events 2026-04-09 → 2026-04-09.
• Practical note: As of 2026-04-09, practitioners evaluating model-weight storage or transfer can treat DMX as a newly announced compression-format project, but should verify implementation details and compatibility in its repository before adopting it.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.