Case study
What Weights.gg got right, and why free could not last
For about three years, Weights.gg was the place most people first made an AI voice cover. It was free, it ran in the browser, and it held the largest community library of RVC voice models anywhere. It closed on 31 March 2026. This is an attempt to understand it honestly: what made it work, what broke, and what a much smaller voice tool has to do differently if it wants to still be here in three years.
Weights.gg won on zero friction and a library that grew itself. It lost on the arithmetic underneath: free GPU generation for a very large, mostly non-paying audience, a library that was both its moat and its legal exposure, and a 2025 attempt to add prices to a habit that had been free for two years. None of that was a failure of the technology. RVC is open source and works exactly as well as it did. The lesson is about the shape of the business, not the model.
What it was
Weights launched in 2023 as a browser tool for AI covers: upload a song, choose a voice from the community library, get the song back in that voice. Over time it added text-to-speech, AI images, character chat, video, and one-click training of your own model. A free account could train several models a week; a premium plan removed the limits and shortened the queue. Longer audio cost credits.
Behind it was a San Francisco team of about six people and roughly four million dollars of venture funding, per PitchBook as cited by The Information. Its second product, Replay, was a free, open-source desktop app that ran covers and training on the user's own GPU while pulling models from the website. The Replay README at the time pointed users to more than thirteen thousand free models on Weights.gg; later write-ups describe the library as hundreds of thousands strong.
Its advantage
Three things, and they compounded.
- No barrier at all. No GPU, no Python, no install, no card. A teenager on a school laptop could make a cover in the time it takes to find the song.
- A library that grew itself. Every model anyone trained became content for the next visitor. This is the classic network effect, and it is why a six-person team reached an audience across five continents. Semrush recorded about 300,000 visits in December 2025 with roughly 95 percent arriving direct: people typed the address in. It was a destination, not a search result. Its biggest countries were the United States, Brazil, South Korea, Vietnam and Turkey.
- Replay moved the cost to the user. Letting heavy users run covers and training on their own hardware, for free, was the smartest thing the company built. It kept the most enthusiastic people engaged without putting them on the company's GPU bill.
It is worth being clear that these were real achievements. Most tools that call themselves a replacement have none of them.
What went wrong, in the order it happened
1. The arithmetic never worked
Every free cover, every free training run, every free image spent GPU time the company paid for. Its farewell post, as reported by several outlets, said the platform had never been profitable, that multiple monetization approaches had not covered operating costs, and that handling abuse and spam took more resources than it could afford. Its audience was enormous and, by geography and age, largely unable or unwilling to pay. Four million dollars buys a lot of GPU time, but not an unlimited amount.
2. The July 2025 changes
In mid-2025 the company tried to fix the arithmetic all at once. According to user reviews on Trustpilot from that period: audio length limits were cut to a few minutes, raw outputs became premium-only, text-to-speech was removed, the character chats were deleted, daily limits were halved, queue caps appeared, and one reviewer describes needing 1,500 credits to create anything. The site's rating settled at 2.9 out of 5. Reviews from before the change praise the best AI site they had used; reviews from after it accuse the company of stealing.
The pattern matters more than the details. A price added to a habit that has been free for two years reads as a betrayal, whatever the price. And the people the free tier had attracted were, by selection, exactly the people least likely to pay. The changes cost the goodwill that made the community valuable without producing the revenue that would have justified it.
One detail from the reviews is telling: a user called the character chats the main thing they came for. A meaningful part of the audience was there for a feature that had nothing to do with voice models, running on the same GPU budget.
3. The library was also the liability
The catalogue that made the site valuable was full of voices nobody had consented to: pop stars, actors, cartoon characters, sitting politicians. Uploads went up first and were moderated later, if at all. Moderation cost scaled with usage; revenue did not. Around the same time, voice-print privacy lawsuits were landing across the industry. The asset and the exposure were the same object.
4. The ending
In January 2026 the company announced “the difficult decision to shut down Weights on March 31st, 2026,” and told users to download their creations and request an export. Petitions to save it collected signatures on Change.org. The site closed on schedule. In May, reports followed that OpenAI had acquired the intellectual property and the six-person team, dispersed the engineers across its own groups, and did not plan to release anything similar. One analysis read the deal less as a hire than as a way to retire a public catalogue of unauthorised voices.
Three months after closing, the address was still receiving tens of thousands of visits a month. The habit outlived the site.
What users actually lost
Read the petitions and the reviews and the loss is not the tool. It is the work. Not every model on Weights.gg could be downloaded; some downloads cost credits; some were never offered. When the site closed, models and covers people had built over years went with it. Anyone who had exported their .pth and .index files lost nothing at all, because those files work in any RVC interface and do not depend on the platform that made them.
If you had models there: search your downloads folder for .pth and .index before assuming the worst. What is and is not recoverable is covered on its own page.
The lessons, and what NiceVois chose
NiceVois is a much smaller thing than Weights.gg was, and that is deliberate. The size is the point. Here is what the case study changed, or confirmed, in how it is built.
| Where Weights.gg ended up | What NiceVois does instead |
|---|---|
| Free GPU generation for everyone, paid for by investors | Your first training is free; after that the price is shown before anything starts. There is no habit to take away later. |
| A public library of other people's voices | No public library. You train from audio you have the right to use, and you confirm that before training starts. |
| Models that could not always be downloaded | Every model is a .pth, .index and ZIP you download. If this site closed tomorrow, you would still have them. |
| Covers, images, chat, video, and training | One workflow: train a voice, hear it work, convert, separate, clean up. Depth in audio rather than breadth across media. |
| Growth first, price later | Small enough to be paid for by the people who use it, with no growth target that forces free. |
Two of those deserve a sentence each.
Ownership is the opposite of what killed Weights.gg. Their moat was a catalogue of voices nobody owned, and that catalogue is what got bought and buried. A model of your own voice, in files on your own drive, is the safest possible version of this technology: for you, and for the tool that made it.
Replay was right. Pushing work to the user's own computer is the honest way to serve heavy users without going broke. NiceVois makes the training easy and gives you the files; if you convert a lot, running your model locally is free, and this site will tell you how.
If you came here from Weights.gg
Plainly: NiceVois replaces one part of Weights.gg, the part with the fewest replacements. It trains a voice from audio and hands you the files. It is not free without limit, and it has no library of celebrity voices. If what you want is free covers with famous voices, a local tool such as Applio is the honest answer, and the alternatives page says so. If what you want is your own voice as a model you keep, that is exactly what this is for.
Upload audio you have the right to use and download a portable .pth, .index and ZIP. Your first training is free.
Train my RVC modelSources: Semrush traffic data, Trustpilot reviews, The Information, Biometric Update, Implicator, a user's account of the shutdown, the Replay repository. Figures such as peak monthly visits vary between sources and are reported as approximate. Corrections welcome at info@nicevois.com.