First: your saved models still work
If you downloaded models from Weights before it shut down, those .pth files and their .index files still work in any RVC tool. Back them up somewhere you control. Our page on the Weights.gg shutdown walks through finding and loading them again.
Where RVC models are shared now
Hugging Face hosts a large number of RVC models. Searching for RVC or the voice you want usually turns up repositories with a .pth file and an index, often zipped together. Quality varies a lot, so read the description and any sample audio before downloading.
The AI Hub community, one of the largest groups around RVC, shares models and keeps public documentation on training and settings. Applio, the open-source RVC app, includes a model search as well. Beyond those, many model makers share their own work through Discord servers and personal pages.
Signs of a good model
A good model page tells you who made it, which RVC version it uses, roughly how much audio and how many epochs went into it, and includes a short audio sample. Files are usually a .pth of around 55 to 60 MB for a v2 model, plus an .index file that can be larger. A .pth that is several hundred megabytes is usually a training checkpoint, not a finished model, and will often fail to load.
Test every new model on the same short line you used for your other voices. Comparing like for like makes it obvious which models are worth keeping and which ones belong in the trash.
What to check before you download
Version and sample rate: RVC v1 and v2 models are not interchangeable in every tool, and you can only merge models of the same version. Index file: models work without one, but a matching .index usually sounds closer to the intended voice. Training data: a short note on how many minutes and epochs went in tells you a lot. Samples: listen before you trust a model with a real project.
Also check whether the model includes its index from the same training run. A mismatched index is a common reason a model sounds off.
The safety risk most people miss
A .pth file is a PyTorch file, and PyTorch's standard format can carry executable code, not just voice weights. Most models are harmless, but loading a file from an unknown source is a real risk. Download from people and places with a track record, never run installers or scripts bundled with a model, and keep your tools up to date.
How to load a model you downloaded
Unzip the download and keep the .pth and its .index together in a folder named after the voice. In the RVC WebUI the model goes in the weights folder and the index is selected at conversion time. In Applio and Xport Studio you add the model through the interface; in Xport, drag the .pth into the Library and it is ready for the vocal converter.
Then run a test with sensible defaults before judging it. Our guide to the best RVC settings lists safe starting values, so a model is not written off because of a wrong transpose or index rate.
Rights and real voices
Many public models imitate real artists. Using them for private experiments is one thing; releasing music with a real singer's voice without permission can break platform rules and laws on likeness and impersonation, and it puts your release at risk. If the goal is music you will put out, use a voice you have the right to use.
Make your own instead
Two ways give you a voice nobody else has. Train one from your own vocals: our guide on how to train an RVC model explains how much audio you need, and Xport Studio's Train Clone runs it on your computer, see voice cloning. Or blend two voices you already have into a new, original one with the Voice Blender, as described in how to merge RVC models.
Either way, keep your models organized. Xport's Library stores every voice with a picture, title, and tags, so a folder of cryptic .pth files becomes something you can actually browse. Voice features are free for 7 days, then part of Xport Pro, $79 one time.