| --- |
| license: apache-2.0 |
| base_model: mlx-community/gemma-4-12B-it-6bit |
| library_name: mlx |
| pipeline_tag: image-text-to-text |
| language: |
| - en |
| - de |
| tags: |
| - mlx |
| - gemma |
| - gemma4 |
| - fine-tuned |
| - dora |
| - lora |
| - ailey |
| - openminded |
| --- |
| |
| # AileyCore-12B |
|
|
| **AileyCore-12B** is a fine-tuned, adapter-merged derivative of **Google Gemma 4 (12B, instruction-tuned)**, |
| optimized to run locally on Apple Silicon via the [MLX](https://github.com/ml-explore/mlx) framework. |
|
|
| It powers **A!ley**, the on-device assistant persona created by **OpenM!nded / Simon van de Loo**. |
|
|
| - **Developed by:** OpenM!nded (Simon van de Loo) |
| - **Model type:** Multimodal (text + image + audio input, text output), decoder-only |
| - **Base model:** [`mlx-community/gemma-4-12B-it-6bit`](https://huggingface.co/mlx-community/gemma-4-12B-it-6bit) (Google Gemma 4 12B-IT, 6-bit quantized) |
| - **License:** Apache License 2.0 |
| - **Languages:** English, German |
| - **Quantization:** 6-bit (q6), preserved through the merge |
|
|
| --- |
|
|
| ## What it is |
|
|
| AileyCore-12B is Gemma 4 12B-IT with a lightweight identity + behavior fine-tune baked directly |
| into the weights. The adaptation was performed with a **mixed DoRA/LoRA** scheme and then **merged** |
| back into the base weights, so no separate adapter is required at inference time. |
|
|
| The identity ("A!ley", created by OpenM!nded / Simon van de Loo) is embedded in the weights and |
| remains stable **with or without** a system prompt. |
|
|
| ## Intended use |
|
|
| - Local, privacy-respecting assistant on Apple Silicon (M-series) Macs |
| - Conversational reasoning, writing, and general assistance in EN/DE |
| - Multimodal understanding (image / audio input) inherited from Gemma 4 |
|
|
| ### Out of scope |
|
|
| - Any use prohibited by applicable law |
| - Safety-critical, medical, legal, or financial decision-making without human oversight |
| - The model can produce inaccurate or biased output; verify important information |
|
|
| --- |
|
|
| ## How to use (MLX) |
|
|
| Because this is a Gemma 4 *unified* (multimodal) checkpoint, load it with **`mlx_vlm`**: |
| |
| ```python |
| from mlx_vlm import load, generate |
| from mlx_vlm.prompt_utils import apply_chat_template |
| |
| model, processor = load("CptShaggy/AileyCore-12B") |
| |
| messages = [{"role": "user", "content": "Wer bist du?"}] |
| prompt = apply_chat_template(processor, model.config, messages) |
| print(generate(model, processor, prompt, max_tokens=256, verbose=True)) |
| ``` |
| |
| > Note: plain `mlx_lm` cannot load the `gemma4_unified` architecture — use `mlx_vlm`. |
| |
| --- |
| |
| ## Training details |
| |
| | Setting | Value | |
| |---|---| |
| | Method | Mixed **DoRA** (attention) + **LoRA** (MLP), merged into base | |
| | DoRA targets | `q_proj`, `v_proj` | |
| | LoRA targets | `gate_proj`, `up_proj`, `down_proj` | |
| | Rank / Alpha | 8 / 16 (scale 2.0) | |
| | Sequence length | 1024 | |
| | Gradient accumulation | 8 | |
| | Learning rate | 1e-4 | |
| | Selected checkpoint | best (val_loss ≈ 1.24) | |
| | Hardware | Apple M4, 24 GB unified memory | |
| | Framework | MLX (`mlx_vlm` + `mlx_lm.tuner`) | |
| |
| The 6-bit quantization of the base model is preserved through the merge; the fused adapter |
| weights are re-quantized to q6. |
| |
| --- |
| |
| ## Limitations & biases |
| |
| Inherited from Gemma 4 plus the fine-tune: the model may produce factually incorrect, |
| outdated, or biased content, and reflects the characteristics of its training data. |
| It is not a knowledge base. Always keep a human in the loop for consequential use. |
| |
| --- |
| |
| ## License & attribution |
| |
| This model is a **Derivative Work** of Google **Gemma 4**, which Google releases under the |
| **Apache License 2.0** (see the official [Gemma 4 license](https://ai.google.dev/gemma/apache_2)). |
| AileyCore-12B is therefore also distributed under **Apache 2.0**. |
| |
| In accordance with Apache 2.0 §4: |
| |
| - The base Gemma 4 weights were **modified** via DoRA/LoRA adaptation and merged. Modified |
| components are noted in `AILEY_MERGE_INFO.json` and this model card. |
| - A copy of the Apache 2.0 license is included (`LICENSE`). |
| - Attribution notices are provided in `NOTICE`. |
| |
| Gemma is a trademark of Google LLC. This project is independent and **not** endorsed by or |
| affiliated with Google. Use of the name "Gemma" here is solely to describe the origin of the base model. |
| |
| ``` |
| Copyright 2026 OpenM!nded / Simon van de Loo |
| Portions © Google LLC (Gemma 4), Apache License 2.0 |
| |
| Licensed under the Apache License, Version 2.0. |
| You may obtain a copy of the License at |
| http://www.apache.org/licenses/LICENSE-2.0 |
| ``` |
| |
| ## Citation |
| |
| ```bibtex |
| @misc{aileycore12b_2026, |
| title = {AileyCore-12B: A Gemma 4 fine-tune for the A!ley assistant}, |
| author = {van de Loo, Simon and OpenM!nded}, |
| year = {2026}, |
| note = {Fine-tuned and merged from Google Gemma 4 12B-IT (Apache 2.0)} |
| } |
| ``` |
| |
