Shotcraft
Voice and speech

Kokoro-82M

A very small open-weights TTS model - 82M parameters, StyleTTS2 architecture with an ISTFTNet vocoder - released under Apache-2.0. Version 1.0 ships 54 voices across 8 language codes at 24kHz output.

A speech model by hexgrad (Kokoro) · read 2026-08-14 · sourcing high

Resolution
24kHz
Weights
Downloadable
API
No
Commercial use
Vendor states commercial use is permitted

It runs fast on modest hardware, which makes it the usual choice for local or embedded TTS.

Commercial use — Vendor states commercial use is permitted

Licence is Apache-2.0, which permits commercial use, modification and redistribution, with attribution and a patent grant, and no share-alike obligation.

The licence notes recorded during research were read for a commercial-use position and did not state one. Read the licence before you rely on any of this. Nothing on this page is legal advice, and this site does not tell you a model is cleared for commercial use.

Specification

Vendor figures, read 2026-08-14Read from hexgrad (Kokoro)'s own documentation on 2026-08-14. Vendors change these between model versions without announcing it, so check theirs before you commit a shot to a number. Sourcing for this entry is strong: the figures came from pages we retrieved and read. Check the vendor's current documentation

What it takes in
text
What it makes
audio
Maximum resolution
24kHz
Controls
speed-control
Weights
Downloadable
API
No
Pricing
Free to self-host under Apache-2.0; compute cost only. The model card notes third-party hosted rates under $1 per 1M characters of input text.
Released or updated
2025-01-27 (v1.0); v0.19 weights released 2024-12-25
Languages
8, not individually listed by the vendor

Strengths

  • Apache-2.0 - the cleanest licence position of any model here
  • 82M parameters runs on CPU or modest GPU, and is cheap at scale
  • 54 voices across 8 language codes, 24kHz output
  • Ranked top of TTS Spaces Arena at release despite training on under 100 hours of audio
  • ONNX build available for easy deployment

Weaknesses

  • No voice cloning - fixed voicepacks only
  • No emotion or style control
  • 8 language codes is limited for localisation
  • No official hosted API; you build the serving layer
  • Quality ceiling below large expressive models for character work

Best for

Where the sources disagree

Recorded rather than resolved. Picking the tidier number would hide the fact that the vendor's own pages do not agree.

What to understand before using it

Explainers naming the mechanisms this model's own documentation mentions.

MCP servers that front hexgrad (Kokoro)

Servers in the media MCP directory whose own description or homepage names this vendor. Absence of a server here is not evidence that none exists: most publish no tool list at all.

See it from the vendor

Shotcraft hosts no sample output. Judging a model means watching its own showreel with your own eyes, and a clip re-hosted here would be a clip we chose.

Nearby in the directory

Azure AI Speech text-to-speech (Neural, Neural HD, MAI voices)

Microsoft's enterprise speech synthesis service, with 200+ prebuilt neural voices across 80+ locales plus tiers for HD voices, multilingual voices, mu

Voice and speech

Sonic 3.5

Cartesia's current TTS model, documented at sub-90ms first-byte latency across 42 languages. It is tuned for conversational delivery and for reading a

Voice and speech

XTTS-v2

An open-weights multilingual TTS and voice cloning model covering 17 languages at 24kHz, still widely used for local voice cloning. Coqui Inc. shut do

Voice and speech

Sources

Read on 2026-08-14. A source that has changed since then will not be reflected here until the next pass.