Tencent Hunyuan HyOCR-1.5 Released: 1B Parameters, Fully Open-Source OCR Stack, 6.37x Inference Speedup
On July 13, 2026, Tencent Hunyuan released HyOCR-1.5 — billed as the first end-to-end OCR expert model to completely open-source its training, inference, and model weights. Despite only 1B parameters, it covers 8+ text-centric tasks and ranks first on OmniDocBench v1.6 with 94.74 points, beating both closed-source and open-source peers, with end-to-end inference at 1.408 seconds per page.
Core performance
| Acceleration backend | Speedup | Notes | |----------------------|---------|-------| | Transformers | 6.37x | HyOCR + DFlash speculative decoding | | vLLM | 2.14x | Throughput doubled at equal accuracy | | End-to-end latency | 1.408 s/page | Measured on OmniDocBench | | OmniDocBench v1.6 | 94.74 | #1 end-to-end |
Technical highlights
- DFlash speculative decoding framework — isolates slow long-document decoding into a standalone acceleration module
- Upgraded training recipe — 4K resolution + 128K context window + RL recipe post-training
- Agentic Data Flow — extends 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, mixed languages
- OCR + Agent is a key combination for 2026 H2 — coding agents need document understanding; HyOCR-1.5 is the first open, industrial-grade, full-stack option in this direction
- Friendly to small developers — training + inference + weights all open, enabling direct domain fine-tuning
- DFlash's engineering pattern is replicable — other OCR / VLM models can adopt it, lifting industry-wide efficiency
- Tencent's "Hy series" ecosystem is forming — Hy3 (Agent LLM) + HyOCR-1.5 appeared the same week; HyVL / HyEmbeddings may follow
- 94.74 over general multimodal models — one of the most persuasive data points for the "specialist > generalist" narrative
- GitHub repo not yet public — as of 7/14 no public Tencent-Hunyuan/HyOCR repository was found; the actual availability of "full-stack open source" needs monitoring
- DFlash compatibility with upstream Transformers / vLLM — the 6.37x benchmark was measured on a specific Transformers version; production integration may require adaptation
- Real accuracy for 331 languages and ancient scripts unpublished — low-resource OCR evaluation details await the code release
- Cost comparison vs commercial OCR APIs — self-hosting HyOCR-1.5 vs calling Baidu / Alibaba OCR APIs requires real-world validation of cost/accuracy trade-offs
Source: Tencent Hunyuan official announcement (WeChat) · released July 13
Analysis
1. "Full-stack open source" as a paradigm shift for OCR
OCR large models have historically taken two paths: closed-source paid APIs (Baidu / Alibaba / Tencent's own OCR services) or "semi-open" weight-only releases. HyOCR-1.5 is the first end-to-end OCR expert model to publish training code, inference code, model weights, and evaluation scripts. This matters greatly for academia and small/mid-size developers, who can fine-tune domain-specific OCR (medical / legal / financial documents) on top of it instead of training from scratch.
2. 1B parameters + 6.37x speedup: small models can do real work
Against the "bigger is better" default, 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 small model drafts tokens that the larger model verifies in one pass — a mature LLM inference technique, but HyOCR-1.5 is the first to deeply integrate it into an OCR setting with 6.37x acceleration.
3. Agentic Data Flow connects OCR to agent workflows
Via Agentic Data Flow, HyOCR-1.5 plugs OCR capability into agent pipelines: 331-language coverage (low-resource OCR such as Tibetan / Burmese), ancient script recognition (oracle bone script, bamboo slips), and multi-image QA. It can serve directly as a "visual perception plugin" for Claude Code, Cursor, or domestic agents — given a PDF, it parses text, tables, and formulas automatically without the agent assembling an OCR pipeline itself.
4. 94.74 on OmniDocBench v1.6 is ceiling-level
OmniDocBench is an authoritative document understanding benchmark covering papers, books, tables, receipts, and formulas. HyOCR-1.5's 94.74 beats general multimodal models (GPT-4V, Gemini, Qwen-VL, InternVL) and OCR specialists (PaddleOCR, MinerU) — a strong case of a specialized small model outperforming general-purpose large models.
5. Sustained contributions to China's LLM "infrastructure layer"
Recent weeks show a steady cadence of infrastructure-level releases: RedKnot (Xiaohongshu, KV cache splitting), LongCat Owl (Meituan, all-ASIC training), Hy3 (Tencent, 295B MoE agent LLM), ForgeTrain (ModelBest, AI-written training framework), and now HyOCR-1.5 — production-grade engineering rather than leaderboard chasing.
6. The "small model punching up" narrative
The official tagline "1B frame, above-class results" echoes 2026's core Chinese LLM narrative of small models beating larger ones on specialized tasks, following VibeThinker-3B (06-28), Qwen3.6-27B (06-29), and LongCat-2.0 (07-05).