gemma-3-12b-it-backbone-4bit-mlx

Gemma 3 12B IT Text-Encoder Backbone (MLX, 4-bit)

MLX 4-bit conversion of the Gemma 3 12B IT text backbone. It serves as the text-conditioning encoder for DramaBox TTS in mlx-speech, exposing the per-layer hidden states the audio model conditions on. It is not a chat or text-generation model.

Backbone, not a full Gemma. The language-model head and generation path are removed. It emits hidden states only. Use a standard Gemma checkpoint for chat or text generation.

Model Details

• Developed by: App Automaton
• Upstream model: google/gemma-3-12b-it, converted from the gemma-3-12b-it-qat-q4_0 release
• Role: text-encoder backbone. Produces all 49 hidden states consumed by the DramaBox prompt pipeline.
• Quantization: MLX 4-bit affine, group_size=64, with bf16 scales and norms
• Runtime: MLX on Apple Silicon

Contents

How to Get Started

This backbone is the text encoder for DramaBox. Download it alongside the DramaBox weights:

hf download appautomaton/gemma-3-12b-it-backbone-4bit-mlx \
  --local-dir models/gemma_3_12b_it_backbone/mlx-4bit

It is then passed to DramaBoxModel.from_dir(..., gemma_dir=...). See the DramaBox card for the full pipeline. To load the backbone directly:

from mlx_speech.models.gemma3_text import load_gemma3_text_model, LTXVGemmaTokenizer

gemma, _ = load_gemma3_text_model("models/gemma_3_12b_it_backbone/mlx-4bit")
tokenizer = LTXVGemmaTokenizer.from_dir("models/gemma_3_12b_it_backbone/mlx-4bit")

Intended Use

Text conditioning for LTX-2-derived audio diffusion models, specifically DramaBox TTS. It is reusable as a Gemma 3 12B feature extractor on Apple Silicon, but it carries no generation head.

Links

• Source code: appautomaton/mlx-speech
• Project page: appautomaton.renocrypt.com/mlx-speech
• Paired model: appautomaton/dramabox-tts-3.3b-bf16-mlx
• More from App Automaton: GitHub · Hugging Face

License

Gemma. Use is governed by Google's Gemma Terms of Use. By downloading or using these weights you agree to those terms.