mlx-vision

Speculative OCR decoding for Apple Silicon

2026 · Python · MLX · Vision-Language Models · Speculative Decoding

mlx-vision

Problem

Modern OCR vision-language models can run locally on Apple Silicon, but autoregressive decoding makes long documents slow and costly to iterate on.

Approach

Built an MLX toolkit with baseline, training-free semi-autoregressive lookup, DSpark-style, DFlash, Eagle3, and MTP paths. It supports DeepSeek-OCR-2, GLM-OCR, and Unlimited-OCR through a CLI, Python API, benchmark runner, and OpenAI-compatible server.

Result

A public, MIT-licensed OCR inference package with target verification for lossless temperature-zero neural speculation, PDF-page input, telemetry, and a content-hash vision cache.

4 OCR presets

MLX Apple Silicon

One MLX toolkit for 4 OCR presets, training-free and neural speculation, with target-verified lossless temperature-zero output.