日本語OCRモデル Sarashina2.2-OCR を MLX へ移植する実装記録This article documents the process of porting the Japanese OCR model…

A developer's write-up on the Zenn platform documents the work of porting Sarashina2.2-OCR, a Japanese-focused optical character recognition model, to MLX, Apple's machine learning framework for its own silicon. The account matters because running capable Japanese OCR locally, rather than through a cloud API, appeals to developers who care about privacy, offline use, cost control, and the ability to process documents without sending them to a third party. The article's particular value lies in filling gaps that the official model card leaves open, turning a partially documented release into something reproducible on a Mac.

Sarashina is a family of Japanese language models developed by SB Intuitions, a SoftBank subsidiary. The lineup has grown from text-only large language models toward multimodal variants, and Sarashina2.2-OCR appears to belong to the vision-capable branch, pairing an image encoder with a language decoder so that scanned pages or photographs can be transcribed into text. Japanese OCR is a demanding task because the writing system mixes kanji, hiragana, katakana, and Latin characters, often in dense layouts with vertical text, furigana, and complex tables. Models tuned specifically for Japanese tend to outperform general-purpose OCR on these cases, which is part of why a local, high-accuracy option is attractive.

MLX is Apple's array framework, introduced in late 2023, designed to take advantage of the unified memory architecture in M-series chips. Because the CPU and GPU share the same memory pool, models can run without the copying overhead common on discrete-GPU systems, and MLX has become a practical route for running language and vision-language models on consumer Mac hardware. The ecosystem around it includes projects such as mlx-lm for text models and mlx-vlm for vision-language models, which provide reference implementations that a port like this one can lean on or extend.

The core of the porting effort, according to the write-up, was reconstructing implementation details that the model card did not spell out. Converting a model from a PyTorch-based release to MLX is rarely a matter of translating weights alone. The author had to investigate how the vision component preprocesses images, including resizing, normalization, and any tiling or patching applied before the encoder, since a mismatch there quietly degrades accuracy without producing an obvious error. Getting the prompt template and special tokens right is another common sticking point, because OCR models often expect a specific arrangement of image placeholders and instruction text, and the tokenizer behavior for Japanese text must be reproduced faithfully.

Beyond preprocessing, porting typically involves mapping each layer of the architecture to MLX equivalents, aligning tensor shapes and weight naming conventions, and verifying that attention, positional encoding, and the connection between the vision encoder and the language model match the original. The article's framing suggests the author worked backward from the reference code and observed outputs to confirm that the MLX version produced results consistent with the source model, which is the usual way to validate a port when documentation is incomplete. This kind of careful cross-checking is what separates a working reproduction from one that merely runs but returns subtly wrong transcriptions.

The broader context is a growing interest in local inference on Apple hardware. As Mac machines ship with increasing amounts of unified memory, developers have been moving models that once required a server or a dedicated GPU onto laptops and desktops. Tools such as llama.cpp, Ollama, and the MLX projects have lowered the barrier, and OCR is a natural fit for local deployment because documents are frequently sensitive. At the same time, the reliance on community porting efforts highlights a recurring issue: model cards for openly released Japanese models sometimes omit the preprocessing and prompting specifics needed to reuse them outside their original framework, leaving practitioners to reverse-engineer the missing pieces.

For readers, the article functions as both a practical guide and a case study in that reverse-engineering process. Anyone attempting a similar port would benefit from its notes on where the official documentation fell short and how the author resolved each ambiguity. It is worth treating the specific findings as tied to this particular model version, since preprocessing conventions and tokenizer details can change between releases, but the general methodology of validating outputs against a reference implementation is likely to remain applicable to future Sarashina models and other Japanese vision-language systems.