| --- |
| pipeline_tag: image-text-to-text |
| language: |
| - multilingual |
| tags: |
| - baidu |
| - vision-language |
| - ocr |
| - custom_code |
| - mlx |
| - mlx-vlm |
| - quantized |
| - affine-quantization |
| - 8-bit precision |
| license: mit |
| library_name: mlx |
| base_model: baidu/Unlimited-OCR |
| --- |
| |
| # Unlimited-OCR — MLX 8bit |
|
|
| This repository contains the **8-bit affine quantized MLX conversion** of Baidu's official [`baidu/Unlimited-OCR`](https://huggingface.co/baidu/Unlimited-OCR) checkpoint. It is published for Apple Silicon inference with [`mlx-vlm`](https://github.com/Blaizzy/mlx-vlm) and maintained for use by [`shuuul/aimd`](https://github.com/shuuul/aimd). |
|
|
| The model architecture, tokenizer, processor configuration, prompts, and MIT license come from the official checkpoint. Refer to the [official model card](https://huggingface.co/baidu/Unlimited-OCR) and [Unlimited-OCR repository](https://github.com/baidu/Unlimited-OCR) for model details, capabilities, limitations, and attribution. |
|
|
| ## Conversion details |
|
|
| | Setting | Value | |
| |---|---| |
| | Source | `baidu/Unlimited-OCR` | |
| | Tool | `mlx-vlm 0.6.8` | |
| | MLX runtime used for conversion | `mlx 0.32.0` | |
| | Output precision | `8bit` | |
| | Quantization | `affine, group size 64` | |
| | Repository size | 3.66 GiB | |
|
|
| The language-model weights use uniform affine quantization with group size 64. mlx-vlm intentionally leaves unsupported or quality-sensitive multimodal modules at higher precision, so total repository size is larger than a theoretical all-layer quantization. |
|
|
| Conversion command: |
|
|
| ```bash |
| mlx_vlm.convert \ |
| --hf-path baidu/Unlimited-OCR \ |
| --mlx-path ./Unlimited-OCR-8bit \ |
| --quantize --q-bits 8 --q-group-size 64 --q-mode affine |
| ``` |
|
|
| ## Use with AIMD |
|
|
| On macOS/Apple Silicon, install AIMD and run OCR directly: |
|
|
| ```bash |
| uv tool install --force "aimd-tool @ git+https://github.com/shuuul/aimd.git@main" |
| aimd scan.pdf --task ocr --model unlimited_ocr_8bit |
| ``` |
|
|
| `unlimited_ocr` defaults to the 4-bit repository. AIMD runs each page in Unlimited-OCR's single-image gundam mode with prompt `document parsing.`, `cropping=True`, `base_size=1024`, `image_size=640`, and a sliding-window no-repeat n-gram guard. See the [AIMD README](https://github.com/shuuul/aimd#ocr-for-scanned-pdfs-and-images) for the complete runtime behavior. |
|
|
| ## Use with mlx-vlm |
|
|
| ```bash |
| pip install -U "mlx-vlm>=0.6.8" |
| python -m mlx_vlm.generate \ |
| --model mlx-community/Unlimited-OCR-8bit \ |
| --image scan.png \ |
| --prompt "document parsing." \ |
| --max-tokens 8192 \ |
| --temperature 0 |
| ``` |
|
|
| Unlimited-OCR support first shipped in `mlx-vlm 0.6.4`; `0.6.8` or newer is recommended to match the conversion environment. |
|
|
| ## Available MLX weights |
|
|
| | Precision | Repository | Size | |
| |---|---|---:| |
| | 4bit | [`mlx-community/Unlimited-OCR-4bit`](https://huggingface.co/mlx-community/Unlimited-OCR-4bit) | 2.29 GiB | |
| | 6bit | [`mlx-community/Unlimited-OCR-6bit`](https://huggingface.co/mlx-community/Unlimited-OCR-6bit) | 2.98 GiB | |
| | 8bit | [`mlx-community/Unlimited-OCR-8bit`](https://huggingface.co/mlx-community/Unlimited-OCR-8bit) | 3.66 GiB | |
| | bf16 | [`mlx-community/Unlimited-OCR-bf16`](https://huggingface.co/mlx-community/Unlimited-OCR-bf16) | 6.22 GiB | |
|
|
| ## License and attribution |
|
|
| The upstream checkpoint is released under the MIT license. This repository is a format conversion and does not modify the upstream model architecture or claim ownership of Baidu's model. |
|
|
