HunyuanOCR-1.5 is Tencent's lightweight, end-to-end OCR-specialized vision-language model that unifies document parsing, text spotting, information extraction, and text-image translation within a single VLM, building on the validated architecture of HunyuanOCR-1.0 (archived under the v1.0/ subfolder) without redesigning the backbone. Its key upgrades target speed and capability: a DFlash speculative-decoding framework uses a lightweight block-diffusion draft model to propose multiple candidate tokens in parallel, verified by the target model in a single pass, significantly cutting decoding latency on long structured outputs like dense documents, tables, and formulas while preserving the original output distribution; it also supports CPU/consumer-GPU/laptop deployment via a GGUF-converted checkpoint and OpenAI-compatible llama-server, including a DFlash-adapted llama.cpp fork. On the training side, an Agentic Data Flow system — where agents handle material search, tool-based verification, and data-pipeline iteration in a closed loop with engineers — targets long-tail capabilities like low-resource and ancient-script OCR, alongside an upgraded recipe extending maximum image resolution to 4K and context window to 128K tokens, with refined SFT data and reinforcement learning across OCR tasks. The model uses the HunYuanVLForConditionalGeneration architecture (requiring transformers ≥5.13.0), supports native transformers, vLLM (both AR and DFlash modes from a single unified uv-based CUDA 13 environment), and llama.cpp inference paths across 12 task types (document parsing, structured/layout parsing, chart/formula/table extraction, and Chinese-English translation variants), and is released under the Tencent Hunyuan Community License Agreement.
