Overview
Tencent Hunyuan released HyOCR-1.5 on July 13, 2026 — described as the first end-to-end OCR expert model to fully open-source its training pipeline, inference code, model weights, and evaluation scripts. The 1B-parameter model covers 8+ text-centric tasks and ranks first in end-to-end document understanding on OmniDocBench v1.6 with 94.74 points, reaching 1.408 seconds per page in end-to-end latency.
Core performance data
| Acceleration backend | Speedup | Notes | |---------|--------|------| | Transformers | 6.37× | HyOCR + DFlash speculative decoding | | vLLM | 2.14× | Doubled throughput at equal accuracy | | End-to-end latency | 1.408 sec/page | Measured on OmniDocBench | | OmniDocBench v1.6 | 94.74 points | First place, end-to-end |
Technical highlights
- DFlash speculative decoding framework — extracted from long-document decoding as an independent acceleration module
- Upgraded training recipe — 4K resolution + 128K context window + RL post-training
- Agentic Data Flow — extended to 331 low-resource languages, ancient script recognition, and multi-image QA
- 8+ text-centric tasks — document understanding, table recognition, formula recognition, receipts, ID cards, ancient books, handwriting, and mixed-language input
- OCR + Agent is a key combination for H2 2026 — agents like Claude Code and Cursor need document understanding; HyOCR-1.5 is the first open, industrial-grade, full-stack option here
- Full-stack open sourcing is highly developer-friendly — fine-tune your own domain OCR without training from scratch
- The DFlash paradigm is replicable — other OCR/VLM models can adopt it, lifting industry-wide efficiency
- A "Hy-series" ecosystem is forming — Hy3 (Agent LLM) + HyOCR-1.5 (OCR expert) appeared in the same week; HyVL, HyEmbeddings, etc. may follow
- 94.74 over general multimodal models — one of the most convincing data points for the "specialized > general" narrative
- GitHub repository not yet public — as of July 14, no public Tencent-Hunyuan/HyOCR repo was found; the actual availability of "full-stack open source" remains to be observed
- DFlash compatibility with upstream Transformers / vLLM — the 6.37× benchmark was run on specific Transformers versions; production integration may require adaptation
- Accuracy for 331 languages and ancient scripts unpublished — low-resource OCR evaluation details await code release
- Cost comparison with commercial OCR APIs — self-hosted HyOCR-1.5 vs. Baidu/Alibaba OCR APIs requires real business validation of cost–accuracy trade-offs
Source: Tencent Hunyuan official WeChat article on HyOCR-1.5
In-Depth Analysis
1. "Full-stack open sourcing" as a paradigm shift for OCR. Previously, OCR large models followed two paths: closed-source APIs, or "half-open" releases that only open-sourced weights. HyOCR-1.5 publicly releases training code, inference code, weights, and evaluation scripts — enabling academic and small/mid-size developers to fine-tune domain-specific OCR (medical, legal, financial receipts) without training from scratch.
2. 1B parameters + 6.37× acceleration = an engineering manifesto that small models can do real work. Against the "bigger is better" assumption, HyOCR-1.5 uses 1B parameters plus DFlash speculative decoding to run faster and more accurately than comparable 7B–13B OCR models. DFlash parallelizes the slow verify stage of autoregressive decoding — a mature LLM inference technique — but HyOCR-1.5 is the first to deeply integrate it into OCR scenarios with a 6.37× speedup.
3. Agentic Data Flow as the OCR + Agent connector. Beyond an OCR engine, the Agentic Data Flow extends OCR into agent workflows: 331 language expansions (e.g., Tibetan, Burmese), ancient script recognition (oracle bone script, bamboo slips), and multi-image QA. HyOCR-1.5 can serve as a visual-perception plugin for agents — parsing text, tables, and formulas from PDFs without the agent assembling its own OCR pipeline.
4. 94.74 points is ceiling-level on OmniDocBench v1.6. OmniDocBench is an authoritative document-understanding benchmark (papers, books, tables, receipts, formulas). HyOCR-1.5's 94.74 surpasses general multimodal models (GPT-4V, Gemini, Qwen-VL, InternVL) and OCR specialists (PaddleOCR, MinerU) — a case of a specialized model outperforming general ones.
5. Continuity of Chinese LLM infrastructure-layer contributions. Recent weeks show a steady rhythm of infrastructure contributions: RedKnot (Xiaohongshu, KV Cache splitting), LongCat Owl (Meituan, full-ASIC training), Hy3 (Tencent, 295B MoE Agent LLM), ForgeTrain (ModelBest), and now HyOCR-1.5 — production-grade engineering rather than benchmark chasing.
6. "A 1B body punching above its weight" as a national-model narrative. The official slogan echoes the core 2026 narrative of small models surpassing larger ones on specialized tasks, previously seen with VibeThinker-3B, Qwen3.6-27B, and LongCat-2.0. HyOCR-1.5 is that narrative applied to the OCR vertical.