Tencent HunyuanOCR 1.0

HunyuanOCR 1.0 was Tencent's entry into specialised document reading, published on 25 November 2025. Its premise was that the traditional OCR stack — a text detector, a separate recogniser, then a layout engine and post-processing rules — could be collapsed into one vision-language model small enough to be practical. The result weighs about one billion parameters (roughly 2 GB of bf16 weights) and covers document parsing, text spotting, information extraction and text-image translation from a single prompt, returning Markdown for prose, HTML for tables and LaTeX for formulas. The architecture is a vision encoder feeding a 24-layer text decoder with a hidden size of 1024 and 16 attention heads. Two limits mark it as the earlier generation: the context window stops at 32,768 tokens and the encoder accepts at most 4,096 image tokens per picture. Both are exactly what the 1.5 release of July 2026 lifted — to 131,072 tokens and 16,384 image tokens — which is why dense multi-page documents were the weak point of this version rather than a matter of accuracy. Tencent did not retire the model when the successor arrived. Version 1.0 stays available in the same Hugging Face repository, archived under a v1.0 subfolder and loadable on demand, so deployments built against it were not broken by the upgrade. The licence is unchanged and matters in Europe: the Tencent Hunyuan Community License states in capitals that it does not apply in the European Union, the United Kingdom or South Korea, so open weights do not translate into a usable licence for organisations based there.

▸Videos

No videos yet.

▸Reviews