Kokoro-82M
A solid pick for self-hosting and full data control.
Overview
Kokoro-82M is an open-weight text-to-speech model with just 82 million parameters, built on a decoder-only StyleTTS2/ISTFTNet architecture with no diffusion step. It covers 8 languages and 54 voices, trained on only a few hundred hours of permissively licensed audio, which caps its expressiveness and language depth compared with proprietary TTS systems trained on far larger corpora. Released under Apache 2.0, it runs comfortably on a laptop, and once self-hosted keeps all voice data on the user's own infrastructure regardless of the vendor's US origin.
Skill profile
Not disclosed
Strengths
- Open-source and self-hostable
Limitations
- API pricing not disclosed
Who is it for
- you want to control cost or self-host
- you have GDPR constraints
Ideal use cases
- Rapid voice prototyping
- Embedded voice assistants
- Automated voice-over generation
- Text-to-speech accessibility
- Local, cloud-free deployments
Access & availability
Key specifications
Privacy
Run it locally
| Quantization | Disk | RAM / VRAM | Typical hardware |
|---|---|---|---|
| Q4 · recommended | 0 GB | 2 GB | Any recent PC/Mac |
| Q8 · balanced | 0.1 GB | 2 GB | Any recent PC/Mac |
| FP16 · max quality | 0.2 GB | 2 GB | Any recent PC/Mac |
Estimates for a moderate context. Long contexts need more RAM (KV cache).
Deploy
Copy-ready commands generated from this card. Adjust context length and GPU count to your hardware.
OpenAI-compatible server for production on NVIDIA GPUs.
pip install vllm
vllm serve hexgrad/Kokoro-82M \
--max-model-len 32768 \
--tensor-parallel-size 1 \
--dtype autoAdvanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | — |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |