Unlimited OCR MLX 8bit

--- license: mit library_name: mlx pipeline_tag: image-text-to-text tags: - mlx - ocr - document-parsing - vision-language - deepseek-ocr - quantized - 8-bit - apple-silicon base_model: baidu/Unlimited-OCR language: - en - zh --- # Unlimited-OCR-MLX-8bit 8-bit MLX quantization of [baidu/Unlimited-OCR](https://huggingface.co/baidu/Unlimited-OCR) -- a long-horizon document OCR / parsing model (DeepEncoder vision + DeepSeek-V2 MoE decoder) -- running natively on Apple Silicon via `mlx-vlm`. ## At a glance | | | |---|---| | **Source** | `baidu/Unlimited-OCR` @ `ee63731b6461c8afcdcc7b15352e7d2ffecc2ead` | | **Architecture** | DeepEncoder (SAM-ViT + CLIP-L) vision + DeepSeek-V2 MoE decoder (12 layers, 64 experts) | | **Format** | MLX (Apple Silicon native), loads via `mlx-vlm`'s `deepseekocr` | | **Quantization** | 8-bit, group-size 64 (9.41 bits/weight effective) | | **Disk size** | **3.7 GB** | ## Verification (2026-07-12, M-series Mac) Converted with `mlx-vlm` 0.3.9 and **document-OCR smoke-tested locally** before publishing -- this pack correctly extracts all fields from a test invoice (number, date, bill-to, line item, amount, status) with bounding-box grounding. ## Sibling variants | Variant | Size | Status | |---|---|---| | **Unlimited-OCR-MLX-8bit** | 3.7 GB | public | | [Unlimited-OCR-MLX-6bit](https://huggingface.co/majentik/Unlimited-OCR-MLX-6bit) | 3.0 GB | public | | [Unlimited-OCR-MLX-4bit](https://huggingface.co/majentik/Unlimited-OCR-MLX-4bit) | 2.3 GB | private — pending quality review | <!-- sibling-table --> ## Usage Requires `mlx-vlm >= 0.3.9`. This repo already carries the small config shim so `mlx-vlm` routes it through its `deepseekocr` implementation (upstream `model_type: unlimited-ocr` -> `deepseekocr`; processor -> `DeepseekVLV2Processor`): ```python from mlx_vlm import load, generate model, processor = load("majentik/Unlimited-OCR-MLX-8bit") out = generate(model, processor, "<image>document parsing.", ["page.png"], max_tokens=2048) print(out) ``` Prompts follow the upstream convention: `<image>document parsing.` for single documents; the model emits text with `<|det|>...<|/det|>` bounding-box grounding. ## Intended use Document OCR, layout parsing, and text extraction (English + Chinese) locally on Apple Silicon. See the [base model card](https://huggingface.co/baidu/Unlimited-OCR) for benchmarks and long-hori...