Kokoro-82M v1.0 🧠 Model Open source

What it is

Kokoro is an open-weight text-to-speech model with only 82 million parameters, published by the developer hexgrad. It is built on the StyleTTS 2 architecture. Version 1.0 (January 2025) is still the current release. It became popular because it sounds far better than its size suggests and runs quickly on a CPU, a laptop or even in a browser.

Key features

• 82M parameters, 24 kHz output, fast on CPU
• 54 preset voices across 8 languages
• Apache-2.0 weights, so commercial use is allowed
• Python package plus community ports (ONNX, JavaScript, browser)
• Hosted by several inference providers at under $1 per million characters

Where you can use it

Download from Hugging Face (hexgrad/Kokoro-82M) or GitHub and run it with the kokoro Python package. You can also use it through hosted inference providers such as fal, Replicate and DeepInfra. We found no creator app in our directory that documents Kokoro by name.

Pricing and rights

The weights are free under Apache-2.0, including commercial use. You pay only for your own compute or a host's per-character fee. Kokoro does no voice cloning, so you cannot copy a real person's voice with it. The trade-off is that you are limited to the preset voices.

Who it is for

Developers, hobbyists and creators on a budget who want free, private, offline narration for videos, audiobooks or apps.

Verdict

Kokoro is a strong open TTS pick for size and licence. It runs almost anywhere and costs nothing. Its expressiveness is well behind ElevenLabs v3, it covers only 8 languages, it has no voice cloning, and it has not been updated since early 2025.