NNiceVois

What happened

Weights.gg shut down: what happened to your voice models

The short answer

Weights.gg was acquired by OpenAI and its services were discontinued on 1 April 2026. If you did not export your .pth and .index files before then, those models are gone and cannot be retrieved from the platform. The underlying RVC technology is unaffected and still open source, so the voice can be trained again from the original audio.

This page exists because most search results for the shutdown recommend text-to-speech products that cannot produce an RVC model file at all. If what you actually want is a portable .pth back, that is a narrower problem with a clearer answer.

Why it closed

OpenAI acquired the company behind Weights.gg and wound down its consumer products rather than continuing them. The platform had hosted a large public catalogue of community-uploaded voice models, including many unauthorised celebrity and artist voices, which carried obvious legal exposure. OpenAI has indicated it does not plan to revive those products.

What is and is not recoverable

What you hadRecoverable?
Model files you downloaded to your own driveYes — they still work in any compatible RVC interface
Models that only ever lived on the platformNo
Community models made by other peopleNo, unless that person republished them elsewhere
Your original source audioYes, if you kept it — and that is all you need to retrain

Check your drive before assuming the worst. A great many people exported a ZIP at some point and forgot. Search your downloads folder for .pth and .index. Those files are self-contained and do not expire.

The gap the shutdown actually left

Weights.gg did several unrelated jobs at once: it hosted a public model catalogue, it produced AI covers in the browser, and it trained models from uploaded audio. Most services now recommended as replacements only address the first two, or replace the whole thing with a text-to-speech voice you cannot export.

The part with the fewest replacements is training. Running RVC locally requires a capable NVIDIA GPU, a Python environment, and a tolerance for notebooks that break. That is the specific gap worth solving.

Retraining a voice you lost

If you still have the source audio, the model can be rebuilt. RVC training does not depend on the old platform in any way — it produces the same standard .pth and .index files that any compatible interface will load. See the dataset preparation guide for what makes a good source recording, and the epoch guide for how long to train.

Train only voices you are permitted to use

The legal pressure that contributed to the shutdown has not gone away. Use your own voice, or a voice you have clear permission to train. Our acceptable use policy sets out what that means here.

Rebuild the model you lost

Upload permitted audio and download a portable .pth, .index, and complete ZIP that you keep.

Train my RVC model