Tencent HunyuanOCR 1.5

HunyuanOCR 1.5 is Tencent's specialised document-reading model: instead of the usual pipeline of a text detector, a recogniser and a layout engine bolted together, one end-to-end vision-language model of just 1.12 billion parameters handles document parsing, text spotting, information extraction and text-image translation. The output is structured directly — Markdown for body text, HTML for tables, LaTeX for formulas — so the model replaces a whole chain of tools rather than one link in it. The vision encoder accepts images up to 4K, and the text side carries a 131,072-token context window, which is what makes long dense documents processable in a single pass. The 1.5 release of 6 July 2026 deliberately did not touch the backbone inherited from HunyuanOCR 1.0. Tencent spent the generation on two things instead. The first is speed: a speculative-decoding scheme called DFlash runs a lightweight block-diffusion draft model that proposes several tokens in parallel, which the main model then verifies in one pass — an answer to the real bottleneck of end-to-end OCR, where a dense table or a page of formulas means very long autoregressive decoding. The second is reach: the model now runs on CPUs, consumer GPUs and laptops through llama.cpp with a GGUF checkpoint, so document processing no longer requires server-grade hardware. Training was rebuilt around what Tencent calls an Agentic Data Flow, an agent-driven data pipeline aimed at long-tail weaknesses — low-resource languages, historical scripts and questions spanning several images at once. One caveat matters for European readers. Despite open weights on Hugging Face, the model is not distributed under a free licence: the Tencent Hunyuan Community License opens with a clause, in capitals, stating that the agreement does not apply in the European Union, the United Kingdom and South Korea. A company in those territories is therefore not offered restrictive terms — it is offered no licence at all. This is the same wording Tencent dropped from its flagship Hy4 line, which moved to plain Apache 2.0; the OCR model still carries it.

▸Videos

No videos yet.

▸Reviews