Update, 9 November. The pull request was merged on 6 November, so the 16 November retirement date is now in Civitai’s live code rather than a proposal. Separately, a creator reported on 8 November that a site admin commented that “Kohya is coming back as part of the new trainer”. We could find no trace of that in Civitai’s public repository, so treat it as a reported comment, not a plan.
On 28 October we wrote that “Kohya is gone” meant one website’s button: Civitai had switched Kohya off in its on-site trainer behind a feature flag, leaving AI Toolkit as the only engine. We noted at the time that we could find no announcement, and that Civitai’s own code was the clearest statement it had made.
The code has spoken again. A pull request opened today in Civitai’s public repository is titled add retirement gate for classic trainer, and it gives a date.
What it does
The classic trainer is set to deactivate on 16 November 2026. After that date, training submissions through it are blocked regardless of user role, and the tests added in the same change check specifically that moderators are blocked as well as regular users. The interface gets a warning before the date and a blocked state after it.
Two caveats. The pull request is open, not merged, so the date can still move. And it is code, not an announcement, though the warning it adds to the interface should become the announcement once it ships.
A Civitai creator spotted it within hours and wrote it up under a title that summarises the reaction: less than two weeks notice. From today that is eleven days.
Where training goes instead
What replaces it is the newer training studio that creators have been describing since September. The classic trainer was already AI Toolkit only after the Kohya change, so the engine does not change on the 16th. The tool around it does.
The creator who found the pull request listed problems with the newer trainer from their own use: a worse layout for tagging images, video uploads that fail every time, AI Toolkit runs that fail often, a single image the site’s rating system flags costing paid Buzz to train, and models held for moderation with no sign of progress. That is one person’s experience over the past month and we have not used either trainer, so it is a report, not a review. It does line up with what several other creators wrote in their own posts in October.
What to do in the next eleven days
If you train on Civitai, the useful work fits in an evening.
Write down the settings for anything you might want to retrain: base model, network size, steps, learning rate, captioning approach. They live in a form that is about to stop accepting submissions.
Download your datasets, images and caption files both. The images are usually still on your own disk. The captions often are not, especially if you tagged on the site.
Finish or abandon any run you have half configured. Something saved in the classic trainer after the 16th is a draft nobody can submit.
And decide where you train next before you need to. The newer studio, a rented GPU, or your own card are all reasonable. The one thing that is not reasonable is discovering on the 17th that a LoRA you planned to update has no settings anywhere you can reach.
We said last week that a configuration you can only see in a hosted form is one you can lose without notice. Eleven days is not much notice, but it is some, and it is more than the Kohya users got.
A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.
