mannix/ kokoro

635 Downloads Updated 4 weeks ago

Kokoro 82M · audio.cpp engine · 54 voices · mp3 at 24 kHz

82m

ollama run mannix/kokoro:82m

curl http://localhost:11434/api/chat \
  -d '{
    "model": "mannix/kokoro:82m",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='mannix/kokoro:82m',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'mannix/kokoro:82m',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Models

Name

1 model

Size / Usage

Context

Input

kokoro:82m

190MB · - context window · Text · 4 weeks ago

kokoro:82m

190MB

-

Text

Readme

Kokoro 82M — text to speech, 54 voices, for xOllama

Kokoro 82M (Q8_0 GGUF) · audio.cpp engine · 54 voices · mp3 at 24 kHz · OpenAI /v1/audio/speech

Runs on xOllama, not on stock ollama. Stock ollama pulls this model but cannot show or run it: it carries only media engines. It needs: - an opencoti engine with the audio.cpp sidecar (dev build 2610030147001/b112 or later, set with XOLLAMA_ENGINE_PATH; the engine pinned by the current xOllama release has no media engines and refuses the model at load); - an xOllama built on or after 2026-10-03. audio.cpp tells a GGUF by its file name, and earlier builds hand it the model under its blob name, which it refuses.

Kokoro is a small, natural-sounding voice model, and SurfSense’s default voice. audio.cpp runs it on the CPU, so it needs no GPU.

Use

xollama pull mannix/kokoro:82m
curl http://localhost:22434/v1/audio/speech \
  -d '{"model": "mannix/kokoro:82m", "input": "The first ship came in at dawn.", "voice": "nova"}' -o speech.mp3
xollama media voices mannix/kokoro:82m   # every voice, and the OpenAI names

Voices

The six OpenAI names map to Kokoro’s own; any of its 54 voices can also be asked for by name. Without a voice the template uses af_heart.

OpenAI	Kokoro
`alloy`	`af_alloy`
`echo`	`am_echo`
`fable`	`bm_fable`
`onyx`	`am_onyx`
`nova`	`af_nova`
`shimmer`	`af_heart`

Measured

On bs2 (opencoti b112, audio.cpp on the CPU, 16 threads), a 21-word sentence:


first request, engine start included	5.1 s
output	7.0 s of mp3, 128 kbps, 24 kHz mono
Whisper large-v3 turbo on the output	word for word, default voice and `nova`

Tags

Tag	Weights	Size
`82m`	`kokoro-82m-q8_0.gguf`	189 MB

Source

GGUF: audio-cpp/audio.cpp-gguf · License: Apache-2.0. Docs: xOllama media.

# Kokoro 82M — text to speech, 54 voices, for xOllama

**Kokoro 82M (Q8_0 GGUF) · audio.cpp engine · 54 voices · mp3 at 24 kHz · OpenAI `/v1/audio/speech`**

> **Runs on [xOllama](https://github.com/mann1x/xollama), not on stock ollama.** Stock ollama pulls this model but cannot show or run it: it carries only media engines. It needs:
> - an [opencoti](https://huggingface.co/ManniX-ITA/opencoti-llamafile) engine with the audio.cpp sidecar (dev build `2610030147001`/b112 or later, set with `XOLLAMA_ENGINE_PATH`; the engine pinned by the current xOllama release has no media engines and refuses the model at load);
> - an xOllama built on or after 2026-10-03. audio.cpp tells a GGUF by its file name, and earlier builds hand it the model under its blob name, which it refuses.

Kokoro is a small, natural-sounding voice model, and SurfSense's default voice. audio.cpp runs it on the CPU, so it needs no GPU.

## Use

```shell
xollama pull mannix/kokoro:82m
curl http://localhost:22434/v1/audio/speech \
  -d '{"model": "mannix/kokoro:82m", "input": "The first ship came in at dawn.", "voice": "nova"}' -o speech.mp3
xollama media voices mannix/kokoro:82m   # every voice, and the OpenAI names
```

## Voices

The six OpenAI names map to Kokoro's own; any of its 54 voices can also be asked for by name. Without a `voice` the template uses **`af_heart`**.

| OpenAI | Kokoro |
|---|---|
| `alloy` | `af_alloy` |
| `echo` | `am_echo` |
| `fable` | `bm_fable` |
| `onyx` | `am_onyx` |
| `nova` | `af_nova` |
| `shimmer` | `af_heart` |

## Measured

On bs2 (opencoti b112, audio.cpp on the CPU, 16 threads), a 21-word sentence:

| | |
|---|---|
| first request, engine start included | 5.1 s |
| output | 7.0 s of mp3, 128 kbps, 24 kHz mono |
| Whisper large-v3 turbo on the output | word for word, default voice and `nova` |

## Tags

| Tag | Weights | Size |
|---|---|---|
| `82m` | `kokoro-82m-q8_0.gguf` | 189 MB |

## Source

GGUF: [audio-cpp/audio.cpp-gguf](https://huggingface.co/audio-cpp/audio.cpp-gguf/tree/main/Kokoro-82M-GGUF) · License: Apache-2.0.
Docs: [xOllama media](https://github.com/mann1x/xollama/blob/dev/docs/xollama/media.mdx).

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)