Tencent HyOCR-1.5 Goes Fully Open (1B Params, 6.37x Faster): Cheap OCR for Coding Agents

A Fully Open End-to-End OCR Model

Tencent's Hunyuan team released HyOCR-1.5, which it describes as the first end-to-end OCR large model to fully open-source its training pipeline, inference code, and model weights together. It is a compact 1 billion parameter specialist covering more than eight text-centric tasks — document parsing, layout, tables, formulas, and more.

The performance headline is a new speculative decoding framework called DFlash, delivering a 6.37x speedup under Transformers and 2.14x under vLLM, reaching end-to-end inference of about 1.408 seconds per page. It supports 4K-resolution inputs and a 128K context window, and uses an "Agentic Data Flow" approach to extend into low-resource OCR scenarios.

Why OCR Cost Matters for Coding Agents

Modern coding agents increasingly ingest non-code inputs: PDF specs, API documentation screenshots, scanned architecture diagrams, error screenshots pasted into a chat. Every one of those has to be turned into text before the model can reason about it. Teams typically do this one of two ways:

• Send the image to a frontier vision model — convenient, but you pay premium per-image and per-token rates, and vision tokens add up fast on multi-page documents
• Run a dedicated OCR step — extract text cheaply first, then feed clean text to your coding model at normal text rates

A free, self-hostable 1B OCR model makes the second path far more attractive. Text extraction moves off your metered API bill entirely and onto hardware you already run.

The Cost Comparison

Suppose your agent processes 5,000 document pages a month. Sending each page as an image to a frontier vision model — with per-image handling plus the output tokens to transcribe it — can easily run into the hundreds of dollars monthly, and it consumes context window you would rather spend on reasoning.

A self-hosted 1B model at ~1.4 seconds per page processes those 5,000 pages in under two hours of GPU time. On a modest cloud GPU, that is a few dollars of compute — and the text you feed downstream is clean, so your coding model spends fewer tokens re-interpreting a fuzzy image. The extraction cost effectively drops toward zero at volume.

Where DFlash Fits

Speculative decoding — the technique behind DFlash — uses a small draft model to propose tokens that the main model verifies in bulk, cutting the number of expensive forward passes. A 6.37x speedup is not just about latency; on self-hosted infrastructure, throughput is cost. Doubling or sextupling pages-per-GPU-hour directly divides your per-page compute bill. This is the same principle driving inference-cost reductions across the industry, now applied to document processing.

The Tradeoffs

Self-hosting is not free of effort:

• Ops overhead — you own deployment, scaling, and monitoring for the OCR service
• Quality validation — a 1B specialist may trail frontier vision models on messy handwriting or unusual layouts; test on your real documents
• Break-even volume — for a handful of pages a month, a hosted vision API is cheaper than standing up any infrastructure

The math favors self-hosting once you cross into steady, high-volume document ingestion — exactly the regime enterprise coding agents hit when they parse specs and tickets all day.

The Takeaway

HyOCR-1.5 continues a clear trend: the boring, high-volume parts of an AI pipeline — OCR, embeddings, classification — are being commoditized by small, fast, open models, while frontier spend concentrates on reasoning. Splitting document extraction out to a cheap local model and reserving your premium coding model for actual reasoning is one of the most reliable ways to cut an agent's bill. Estimate the impact on your own workflow with our AI coding cost calculator.