Is XTTS v2 / Coqui TTS Free for Commercial Use? (2026)

No. The XTTS v2 model weights are released under the Coqui Public Model License (CPML) 1.0.0, which explicitly permits only non-commercial use of the model and its audio output — and because Coqui Inc. shut down in January 2024, there is no longer anyone to sell you a commercial license. The confusing part is that the code (the Coqui TTS Python library) is MPL 2.0 and is fine for commercial use — it is the trained weights that are locked to non-commercial. If you need TTS or voice output in a paid product, SaaS, ad-supported video, or client work, the genuinely safe open picks are Kokoro (Apache-2.0), Chatterbox (MIT), and Piper (MIT).

This question comes up constantly because XTTS v2 sounds excellent and clones voices from a few seconds of audio, so people assume "open source on Hugging Face" means "free to ship." It does not. License and code license are two different things, and TTS is one of the worst categories for this trap. This guide separates the two cleanly, gives you a verified license table, and points you at models that actually let you make money.

Is XTTS v2 free for commercial use?

Short answer: no. The XTTS-v2 model card on Hugging Face ships a LICENSE.txt that is the Coqui Public Model License 1.0.0. The license defines what you may do as "non-commercial purposes" and states the restriction directly:

"Non-commercial purposes include any of the following uses of the model or its output, but only so far as you do not receive any direct or indirect payment arising from the use of the model or its output."

It goes further and blocks using the model to bootstrap a commercial product: "Use of the model to train other models for commercial use is not a non-commercial purpose." So you cannot use XTTS v2 to generate a dataset and then claim the downstream model is clean.

Back in 2023, Coqui did sell commercial XTTS licenses (there was a tier around 365 USD/year for companies under 1M USD in revenue or funding). That path is gone. Coqui the company shut down — the founders announced it in late 2023 and services went offline in January 2024. There is no sales team, no portal, and no one to issue or honor a commercial license today. Practically, that makes XTTS v2 non-commercial-only with no legal escape hatch. You can read the exact terms on the XTTS-v2 LICENSE.txt.

The code is MPL 2.0 — so why can't I use it commercially?

This is the single most common point of confusion, so it's worth being precise. There are two separate licenses in play:

• The Coqui TTS library (the Python code) is licensed under MPL 2.0. MPL 2.0 is a permissive, business-friendly license — commercial use is fine. The code is still maintained by the community; an active maintained fork lives at idiap/coqui-ai-TTS (installable as the coqui-tts package).
• The XTTS v2 weights (the trained model) are licensed under the Coqui Public Model License (CPML) — non-commercial only.

So "the code is open" is true and "the model is free for my product" is false, at the same time. You can ship a commercial app built on the MPL-2.0 TTS library, but you cannot ship it loaded with the CPML XTTS v2 weights. To use that library commercially you would pair it with weights that carry a permissive license — which is exactly what the table below is for. For a setup walkthrough of the model itself (non-commercial), see our Coqui TTS / XTTS-v2 local setup guide.

Local TTS license comparison (2026)

Here is the verified license status for the local TTS and voice-cloning models people actually run in 2026. "Commercial OK" means the model weights themselves carry a license that allows paid/commercial use — not just the code.

Two entries deserve a flag because they trip people up. F5-TTS is frequently called "commercial-friendly" because its code is MIT — but the published weights are CC-BY-NC-4.0 (the maintainers say this is due to the Emilia in-the-wild dataset), so the official checkpoints are non-commercial. You would have to retrain on your own commercially-licensed data to ship F5-TTS in a product. And Piper is MIT for the engine, but the voices are a mix — most are permissive, but a few were trained on datasets with their own terms, so confirm the specific voice you ship.

What should I use instead of XTTS v2 for a commercial product?

Pick by what matters most for your use case:

• Want the easiest clean license + low compute? Use Kokoro (Apache-2.0). It is an 82M-parameter model, the weights are about 327 MB, and it runs fast on a plain CPU. Apache-2.0 is about as permissive as it gets — ship it in a SaaS, an app, ad-supported videos, anything. The trade-off is no zero-shot voice cloning; you get a fixed set of high-quality voices (54 voices across 8 languages in v1.0). Full walkthrough in our Kokoro TTS local setup guide.
• Want voice cloning under a clean license? Use Chatterbox (MIT). Resemble AI released it under MIT in 2025; it is a ~0.5B model with zero-shot cloning from a short reference clip plus emotion-exaggeration control, and it embeds a watermark on output. MIT means commercial use with no royalties or revenue share.
• Want CPU-only, embedded, or Raspberry Pi narration? Use Piper (MIT). It is the lightest of the three and built for offline, low-power devices — just confirm your chosen voice pack's license before shipping.

If you specifically came for XTTS-style cloning, Chatterbox is the closest like-for-like replacement that you can actually monetize. For a broader comparison of cloning models, see our local AI voice cloning roundup and the dedicated F5-TTS setup guide (great for personal/research use, just not commercial off the shelf).

What to expect when you run these locally

Seconds-per-sentence figures are useless across machines, so instead here is the mechanism that decides how each of these feels. Local TTS cost comes down to two things: how big the model is (every parameter has to be read out of memory to produce audio), and whether the decoder is autoregressive — generating one token at a time, which cannot be parallelised — or a single feed-forward pass.

• Kokoro is the one that genuinely doesn't need a GPU. It is an 82M-parameter model at roughly 327 MB, so the entire thing sits comfortably in ordinary system RAM and the per-sentence compute is small enough that a plain CPU keeps up. That footprint plus Apache-2.0 is exactly why it is the default commercial recommendation here.
• XTTS v2 and Chatterbox are an order of magnitude larger and want a GPU to feel responsive. Chatterbox is a ~0.5B model, and XTTS v2 follows the Tortoise design — a GPT-style autoregressive stage feeding a neural vocoder — so speech is produced token by token. Sequential decoding is precisely the workload a GPU accelerates; on CPU these still run, they just stop feeling interactive.
• Piper sits at the opposite end by design. It is built for offline, low-power hardware, which is why it is the sensible pick for a Raspberry Pi or an embedded device rather than a workstation.

If you need real latency numbers, generate a handful of your own sentences on your own hardware and time them — the ratio moves with sentence length, sample rate, batch size and quantization. The licensing point holds regardless of what those numbers say: license, not raw quality, is the deciding factor for a paid product, and the permissive models are good enough that staying legal costs you very little.

Key Takeaways

1. XTTS v2 is not free for commercial use. Its weights are under the Coqui Public Model License (non-commercial), and Coqui shut down in January 2024, so no commercial license can be bought.
2. Code license is not model license. The Coqui TTS library is MPL 2.0 (commercial OK), but the XTTS v2 weights are CPML (non-commercial). Both are true at once.
3. F5-TTS has the same split. MIT code, but CC-BY-NC-4.0 weights — the official checkpoints are non-commercial.
4. The safe commercial picks are Kokoro (Apache-2.0), Chatterbox (MIT), and Piper (MIT). Kokoro for the cleanest license and CPU use, Chatterbox for cloning, Piper for embedded/low-power.
5. Always verify the weights license on the model card, and for Piper verify the individual voice pack — open source on Hugging Face does not mean commercial-free.

Next Steps

• Need cloning you can sell? Read the local AI voice cloning comparison to see Chatterbox and OpenVoice v2 side by side.
• Want the no-GPU, Apache-2.0 default? Follow the Kokoro TTS local setup guide.
• Still want to try XTTS v2 for personal/research use? Our Coqui TTS setup guide covers install and voice cloning (non-commercial only).
• Comparing the newest cloning model? See the F5-TTS setup guide — excellent quality, but check the weights license before any paid use.