Kokoro-82M
Open Weights
hexgrad
An Apache-licensed text-to-speech model at just 82 million parameters, covering 8 languages and 54 voices. Built on StyleTTS 2 with an ISTFTNet vocoder and a decoder-only design - no diffusion and no encoder release - it runs comfortably on CPU while delivering quality its maintainers describe as comparable to far larger models. It was trained on a few hundred hours of permissive, non-copyrighted audio for roughly $1,000 in total compute.
Strengths
- Only 82M parameters, so it runs on CPU without a GPU
- Apache-2.0 licensed with no usage restrictions
- 8 languages and 54 voices in v1.0
- Decoder-only StyleTTS 2 architecture with an ISTFTNet vocoder - no diffusion step
- Trained on permissive, non-copyrighted audio, which matters for commercial use
- Roughly $1,000 total training cost, a useful datapoint on small-model economics
Caveats
- Text-to-speech only - it does not transcribe or converse
- No voice cloning; you choose from the 54 packaged voices
- The model card publishes no formal benchmark scores, so none are recorded here
- 8 languages is narrow next to multilingual TTS models covering 20+
- Vendor
- hexgrad
- License
- Apache-2.0
- Release Date
- 2025-01-27
- Modalities
- textaudio
Capabilities
Vision
Audio
Video
Tool Use
Resources
Pricing
Open weights under Apache-2.0 - self-host at no licence cost. At 82M parameters it runs on CPU, so it needs no GPU budget at all.