Instructions to use anquachdev/connect-hibiki-zero-vi with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use anquachdev/connect-hibiki-zero-vi with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] hf download anquachdev/connect-hibiki-zero-vi --local-dir connect-hibiki-zero-vi
- Notebooks
- Google Colab
- Kaggle
- Local Apps Settings
- LM Studio
- Atomic Chat
Connect: Hibiki-Zero Vietnamese and English, built for devices
These are the device builds of the two-way model release/twoway_grpo1400 of huybik/hibiki-zero-vi-full-sft, a fine-tune of Kyutai's Hibiki-Zero 3B. One model translates Vietnamese speech to English and English speech to Vietnamese, while the speaker talks.
The code that runs these files, the apps, and the benchmark results are in anthoai97/onevoice-ai-challenge.
| Folder | Device | Contents |
|---|---|---|
mac/ |
Apple silicon Mac | mlx-q8.safetensors, the model with 8-bit weights for MLX (3.59 GB) |
base/ |
Mac | The Mimi codec, the tokenizer, and the model configuration |
snapdragon/<chip>/ |
Snapdragon NPU | 9 QNN context binaries in ONNX wrappers, 8-bit weights and 16-bit activations, compiled with QAIRT 2.50. build.json lists the AI Hub job IDs. |
snapdragon/host/ |
Snapdragon CPU | Embedding tables, the vocabulary, and the precomputed start of the stream. The same for every chip. |
packages/connect-macos-v1.0.zip |
Apple silicon Mac | The offline Mac app: code, the 8-bit model, the codec, a standalone Python, and every Python package. Unzip it and run bash install.sh; no network needed. |
snapdragon/ folder |
Chip | Tested on |
|---|---|---|
8-elite |
SM8750 | Snapdragon 8 Elite reference phone, Android 16 |
8-elite-gen5 |
SM8850 | Snapdragon 8 Elite Gen 5 reference phone, Android 16 |
8-gen3 |
SM8650 | Snapdragon 8 Gen 3 reference phone, Android 14 |
x-elite |
X1E80100 | Snapdragon X Elite laptop, Windows 11 |
x2-elite |
X2E88100 | Snapdragon X2 Elite laptop, Windows 11 |
Download one device's files:
hf download anquachdev/connect-hibiki-zero-vi --include "mac/*" "base/*" --local-dir weights-release
hf download anquachdev/connect-hibiki-zero-vi --include "snapdragon/8-elite/*" "snapdragon/host/*" --local-dir weights-release
Licence
The weights carry the licence of the base model, CC BY-NC-SA 4.0, which allows non-commercial use with attribution, under the same licence. The Mimi codec in base/ is from Kyutai.
Hardware compatibility
Log In to add your hardware
Quantized
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support
Model tree for anquachdev/connect-hibiki-zero-vi
Unable to build the model tree, the base model loops to the model itself. Learn more.