Instructions to use darkmaniac7/tokforge-assets with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- llama.cpp
How to use darkmaniac7/tokforge-assets with llama.cpp:
Install (macOS, Linux)
curl -LsSf https://llama.app/install.sh | sh # Start a local OpenAI-compatible embeddings server: llama serve -hf darkmaniac7/tokforge-assets:Q4_K_M --embeddings
Install from WinGet (Windows)
winget install llama.cpp # Start a local OpenAI-compatible embeddings server: llama serve -hf darkmaniac7/tokforge-assets:Q4_K_M --embeddings
Build from source code
git clone https://github.com/ggml-org/llama.cpp.git cd llama.cpp cmake -B build cmake --build build -j --target llama-app # Start a local OpenAI-compatible embeddings server: ./build/bin/llama serve -hf darkmaniac7/tokforge-assets:Q4_K_M --embeddings
- Llama new
- LM Studio
- Jan
- Ollama
How to use darkmaniac7/tokforge-assets with Ollama:
ollama run hf.co/darkmaniac7/tokforge-assets:Q4_K_M
- Unsloth Desktop
- Docker Model Runner
How to use darkmaniac7/tokforge-assets with Docker Model Runner:
docker model run hf.co/darkmaniac7/tokforge-assets:Q4_K_M
- Lemonade
How to use darkmaniac7/tokforge-assets with Lemonade:
Pull the model
# Download Lemonade from https://lemonade-server.ai/ lemonade pull darkmaniac7/tokforge-assets:Q4_K_M
Run and chat with the model
lemonade run user.tokforge-assets-Q4_K_M
List all available models
lemonade list
- Atomic Chat
TokForge assets (ONNX, GGUF)
This repo holds the voice and speech files TokForge downloads to read replies aloud, clone a voice and transcribe audio on your phone.
Published by TokForge for on-device inference.
| What it is | Voice and speech assets (text-to-speech, voice cloning, speech recognition) |
| Format | ONNX files for sherpa-onnx (fp32, fp16 and int8), plus one GGUF file (Q4_K_M) |
| Runs on | Android phones (CPU, sherpa-onnx); TokForge for iOS downloads the same voice and speech files |
| Phone class | No catalog class; each part loads on its own and needs more free RAM than its files (the largest, Kokoro's full model, is 346 MB) |
| Download | 1.32 GB (729 files) |
| Base model | hexgrad/Kokoro-82M, k2-fsa/ZipVoice, charactr/vocos-mel-24khz, openai/whisper-tiny.en, BAAI/bge-small-en-v1.5 |
| License | Mixed: ZipVoice under CC-BY-NC (non-commercial), eSpeak NG data under GPL-3.0, the rest Apache-2.0 or MIT (see License) |
TokForge
- Website: https://tokforge.ai
- Discord: https://discord.gg/Acv3CBtfVm
- Google Play: https://play.google.com/store/apps/details?id=dev.tokforge
- iOS TestFlight: https://testflight.apple.com/join/jnufjzRr
Runs on-device in the TokForge app.
Use it in TokForge
You do not download this repo yourself. TokForge fetches only the files a feature needs, when you turn that feature on.
- Install TokForge from Google Play, or join the iOS TestFlight.
- To have replies read aloud, open Settings > Voice & sound and download the Kokoro voice. Standard fetches the int8 model (
model.int8.onnx), Premium the full model (model.onnx); both also fetchvoices.bin,tokens.txtand the English eSpeak NG files fromkokoro-en-v0_19/. - To clone a voice, open Settings > Voice & sound and set up voice cloning. TokForge downloads
encoder.int8.onnx,decoder.fp16.v2.onnx,vocos_24khz.onnx,tokens.txt,lexicon.txtand the English eSpeak NG files fromzipvoice-distill-int8/. - To transcribe audio or the sound of a video, TokForge uses its speech-to-text helper (listed under Settings > Models & Modules). It downloads
tiny.en-encoder.int8.onnx,tiny.en-decoder.int8.onnxandtiny.en-tokens.txtfromsherpa-onnx-whisper-tiny.en/at a fixed commit of this repo, either with the other helper models or the first time you transcribe something. - Everything runs on the phone; once the files are downloaded, these features need no connection.
TokForge for iOS downloads the same Kokoro, ZipVoice and Whisper files.
Collections: TokForge Voice and Helper Models.
Files
| File | Size | What it is |
|---|---|---|
bge-small/bge-small-en-v1.5-q4_k_m.gguf |
24.8 MB | bge-small-en-v1.5 text embeddings (GGUF Q4_K_M); not downloaded by current TokForge versions |
kokoro-en-v0_19/ |
504 MB in 360 files | Kokoro v0.19 text-to-speech: fp32 and int8 models, voices.bin, tokens.txt, eSpeak NG data (GPL-3.0), SHA256SUMS.txt |
sherpa-onnx-whisper-tiny.en/ |
104 MB in 6 files | Whisper tiny.en speech recognition (int8 encoder and decoder, tokens), SOURCE.json, LICENSE.whisper, README.md |
zipvoice-distill-int8/ |
684 MB in 362 files | ZipVoice-Distill voice cloning (CC-BY-NC): encoder, decoders, Vocos vocoder, tokens, lexicon, eSpeak NG data (GPL-3.0) |
Folders with many files are shown as one row; the next section lists what each folder holds.
What is in each folder
| Folder | What it holds | What TokForge downloads from it |
|---|---|---|
kokoro-en-v0_19/ |
Kokoro v0.19 text-to-speech for sherpa-onnx: the full model model.onnx (fp32), the smaller model.int8.onnx, the voice table voices.bin, tokens.txt, eSpeak NG data in espeak-ng-data/ and SHA256SUMS.txt |
One of the two models, voices.bin, tokens.txt and the English eSpeak NG files with their voice variants |
zipvoice-distill-int8/ |
ZipVoice-Distill voice cloning for sherpa-onnx: encoder.int8.onnx, three versions of the decoder (decoder.int8.onnx, decoder.fp16.onnx, decoder.fp16.v2.onnx), the Vocos vocoder vocos_24khz.onnx, tokens.txt, lexicon.txt and its own copy of espeak-ng-data/ |
encoder.int8.onnx, decoder.fp16.v2.onnx, vocos_24khz.onnx, tokens.txt, lexicon.txt and the English eSpeak NG files |
sherpa-onnx-whisper-tiny.en/ |
Whisper tiny.en speech recognition: the int8 encoder and decoder, tiny.en-tokens.txt, SOURCE.json (source commit and sha256 of each file), LICENSE.whisper and a short README.md |
The encoder, the decoder and the tokens |
bge-small/ |
bge-small-en-v1.5 text embeddings as one GGUF file (Q4_K_M) for llama.cpp | Nothing in current versions; TokForge's search embeddings now come from a separate download |
How it was made
- Kokoro: the sherpa-onnx export of Kokoro v0.19 (fp32 and int8), with the eSpeak NG data that comes with it. TokForge serves these files from this repo so voice downloads do not depend on the upstream repo.
- ZipVoice:
encoder.int8.onnx,decoder.int8.onnx,vocos_24khz.onnx,tokens.txt,lexicon.txtandespeak-ng-data/come from k2-fsa's sherpa-onnx release of ZipVoice-Distill.decoder.fp16.onnxanddecoder.fp16.v2.onnxare TokForge's fp16 re-exports of the same decoder weights; v2 only changes the ONNX Runtime session settings stored in the file. Current TokForge versions downloaddecoder.fp16.v2.onnx. - Whisper tiny.en: an unmodified sherpa-onnx int8 export from csukuangfj/sherpa-onnx-whisper-tiny.en at commit
d026532c022f;SOURCE.jsonlists the sha256 of each file. - bge-small-en-v1.5: quantized to Q4_K_M with llama.cpp.
Limitations
- Runtime files for TokForge and sherpa-onnx, not a chat model and not training checkpoints.
- No Hugging Face Inference widget is expected here.
- Kokoro, Whisper tiny.en and bge-small here are English models. ZipVoice's lexicon covers Chinese and English; TokForge uses it for English.
- Contents may evolve as TokForge asset packaging changes.
License
The folders here are under different licenses. zipvoice-distill-int8/ holds k2-fsa's ZipVoice-Distill model, which k2-fsa publishes under CC-BY-NC (non-commercial) on its model card; the repo metadata declares that, the most restrictive license here. The Vocos vocoder in the same folder is MIT. kokoro-en-v0_19/ is Apache-2.0 except its espeak-ng-data/, which is eSpeak NG data under GPL-3.0. Whisper tiny.en and bge-small-en-v1.5 are MIT. Each part is listed under Credits.
zipvoice-distill-int8/espeak-ng-data/ is an identical copy of the same eSpeak NG data, also under GPL-3.0. TokForge's fp16 versions of the ZipVoice-Distill decoder (decoder.fp16.onnx, decoder.fp16.v2.onnx) are the same pre-trained model in another precision and stay under CC-BY-NC (non-commercial).
Credits
- Kokoro-82M v0.19 by hexgrad (Hugging Face). Used under the Apache-2.0 license. In
kokoro-en-v0_19/. - eSpeak NG data (GitHub), in
kokoro-en-v0_19/espeak-ng-data/as shipped with the sherpa-onnx Kokoro export. Used under the GPL-3.0 license. - The Kokoro ONNX export is from sherpa-onnx by k2-fsa (Apache-2.0), via csukuangfj/kokoro-en-v0_19.
- Whisper tiny.en by OpenAI (Hugging Face), in
sherpa-onnx-whisper-tiny.en/: unmodified sherpa-onnx INT8 export from csukuangfj/sherpa-onnx-whisper-tiny.en. Used under the MIT license (LICENSE.whisper). - ZipVoice-Distill by k2-fsa (Hugging Face, GitHub), in
zipvoice-distill-int8/. As k2-fsa publishes it: the code is under the Apache-2.0 license, and the pre-trained model is under CC-BY-NC (non-commercial) because of its Emilia training data. - Vocos mel 24 kHz vocoder by Charactr (Hugging Face). Used under the MIT license. As
zipvoice-distill-int8/vocos_24khz.onnx. - bge-small-en-v1.5 by BAAI (Hugging Face). Used under the MIT license. As
bge-small/bge-small-en-v1.5-q4_k_m.gguf, quantized with llama.cpp. - eSpeak NG data (GitHub), an identical copy in
zipvoice-distill-int8/espeak-ng-data/for the ZipVoice voice. Used under the GPL-3.0 license.
- Downloads last month
- 30
4-bit