The latest commit in atomicbot-ai/atomic-chat is PR #398, merged at 19:56 UTC on October 30, 2026, a change that pins diffusion model files on a per-repository commit basis. The repository description states that this project is a local AI app and inference engine for agents, and that it aims to run open-weight LLMs offline on a personal computer.
The starting point of the change was a file swap upstream. When Unsloth replaced the Z-Image Turbo quantized file on main, the size of the Q4_K_M file changed from 5,017,613,376 bytes to 3,850,130,496 bytes, but the config file still carried the old size and hash. The size and SHA-256 validation rejected the new file, and the decision document records that this reaction was the result of validation doing its job. Until then, clients did not read the revision value of diffusion entries.
So an optional transformer.revision and a revision field for sidecar files were added to diffusion schema version 1, and entries were changed to be interpreted against commits in their own repository. A pin attached to the transformer is not inherited by the VAE or text encoder repositories; main is used as before only when there is no pin, and invalid pins are rejected. Z-Image Turbo's size and hash were updated against Unsloth commit b40f712c569ecfff42f2307dcc43399ae1d71318, and tests covering this were added as well.
However, this cleanup is a change in client behavior. The decision document notes that existing clients ignore the new fields, so a new release is required for the pinning to actually take effect. Size and SHA-256 validation remain mandatory, the metadata was cross-checked against the hub, but it was not verified that images are actually generated with the replaced quantized file, and the core runtime and download protocol are unchanged. File pinning reduces reproducibility problems, but whether the output changes has not yet been confirmed, so a release and an actual generation check are both needed.