Tencent-Hunyuan/HunyuanOCR

What it solves

HunyuanOCR-1.5 is a lightweight, end-to-end vision-language model (VLM) designed to unify various text-centric visual tasks. It solves the problem of slow inference speeds for long structured outputs (like tables and formulas) and the lack of specialized capabilities for low-resource, ancient-script OCR and multi-image text-centric QA.

How it works

The project employs a lightweight VLM architecture and introduces several key optimizations:

• DFlash Speculative Decoding: Uses a lightweight block-diffusion draft model to predict multiple candidate tokens in parallel, which are then verified by the target model in a single pass to reduce latency.
• Agentic Data Flow: An agent-driven system that identifies model weaknesses and automatically generates targeted training data for long-tail capabilities.
• Flexible Deployment: Supports high-performance serving via vLLM and native transformers, as well as consumer-grade deployment on CPUs and laptops via llama.cpp (GGUF format).
• Upgraded Training: Extends image resolution to 4K and the context window to 128K, incorporating reinforcement learning (RL) to refine OCR tasks.

Who it’s for

This tool is for developers and researchers looking for a high-performance, lightweight OCR model that can be deployed on both server-grade GPUs and consumer hardware, and those needing specialized OCR for complex documents, ancient scripts, or multi-image analysis.

Highlights

• Unified OCR Framework: Combines document parsing, text spotting, information extraction, and text-image translation into one model.
• Lossless Acceleration: DFlash speculative decoding speeds up long structured outputs without changing the output distribution.
• Broad Hardware Support: Compatible with vLLM, transformers, and llama.cpp for versatile deployment.
• High-Resolution Support: Handles images up to 4K resolution with a 128K context window.
• Open-Source Pipeline: Provides full SFT and DFlash training pipelines for community extension.