Instructions to use assassindesign/gemma-4-e4b-8bit-fixed with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MLX
How to use assassindesign/gemma-4-e4b-8bit-fixed with MLX:
# Download the model from the Hub pip install huggingface_hub[hf_xet] huggingface-cli download --local-dir gemma-4-e4b-8bit-fixed assassindesign/gemma-4-e4b-8bit-fixed
- Notebooks
- Google Colab
- Kaggle
- Local Apps
- LM Studio
Gemma-4-E4B-it 8-bit MLX (Audio Supported / Fixed)
本模型是将官方原版 google/gemma-4-e4b-it 转换为 Apple Silicon (Mac) 专用的 MLX 格式的 8-bit 量化版本。
该模型由 mlx-vlm 版本 0.5.0 转换生成。有关模型的架构、训练数据和预设能力的更多详细信息,请参阅原始模型卡。
This model is an 8-bit quantized MLX version of the original google/gemma-4-e4b-it, optimized for Apple Silicon Macs. It was converted using mlx-vlm version 0.5.0. For more details about the model architecture and training data, please refer to the original model card.
🌟 此版本的特别说明 (Important Fixes)
早期的 mlx-vlm 版本(如 0.4.x)在转换和量化 Gemma 4 时,存在破坏音频塔(Audio Tower)权重以及丢失 feature_extractor 的 Bug,导致模型在处理音频时出现“失聪”或无限输出网址乱码(如 123456...)的问题。
本仓库的权重使用了最新的 mlx-vlm v0.5.0 重新转换,完美跳过了音频塔的损坏,保留了无损的音频特征提取能力! 现在你可以直接在 Mac 上极速运行 Gemma 4 的原生语音识别(ASR)和音频对话功能。
💻 安装与依赖 (Requirements)
请确保你安装了最新版的 mlx-vlm(需要包含音频修复补丁):
pip install -U "mlx-vlm>=0.5.0" transformers
- Downloads last month
- 377
Model size
3B params
Tensor type
BF16
·
U32 ·
Hardware compatibility
Log In to add your hardware
8-bit
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support