HunyuanOCR (curateclick.com)

Tencent’s HunyuanOCR is an end-to-end, OCR-focused vision‑language model that packs detection, recognition, parsing, IE, subtitle extraction and image translation into a single 1B‑parameter multimodal architecture. The model emphasizes a “single‑prompt → single‑inference” workflow to cut pipeline latency and cascading errors, supports 100+ languages, and can output structured formats (coordinates, LaTeX, HTML, Mermaid, Markdown, JSON) for direct downstream use. In Tencent’s in‑house benchmarks HunyuanOCR leads across tasks: text spotting overall ~70.9 (vs. PaddleOCR 53.4), document parsing Omni score ~94.1, and information‑extraction metrics (cards/receipts/subtitles) all ~92–93, demonstrating consistent gains over both modular OCR stacks and large general VLMs. The release focuses on production readiness: recommended deployment uses vLLM (better throughput/latency today) with sampling params like temperature=0 and max_tokens=16384, while a Transformers path (HunYuanVLForConditionalGeneration) is available but currently trails. System requirements call for Linux, Python 3.12+, CUDA 12.8, PyTorch 2.7.1 and an ~80GB CUDA GPU for 16K‑token decoding; README defaults to bfloat16 and device_map="auto" (watch multi‑GPU sharding). Practical guidance includes prompt patterns (explicit JSON/field enumeration, language constraints), post‑processing helpers (substring dedupe, schema validation), and guardrails to prevent malformed outputs—making HunyuanOCR a compelling, low‑cost single‑model option for multilingual, multi‑format OCR in production.