Tencent's HunyuanOCR-1.5 Claims 6.37x OCR Inference Speedup

Tencent's Hunyuan Vision team has posted an arxiv writeup of HunyuanOCR-1.5, and the interesting bit is not that it hits the top of another leaderboard, it is what they did to get inference fast enough to be worth deploying. The team reports a 6.37x Transformer inference speedup and a 2.14x speedup under vLLM, achieved by adapting a technique they call DFlash to OCR decoding.

Speculative decoding is not new, but OCR is a particularly good fit for it: end-to-end OCR spends most of its time in long autoregressive output, on tables, dense pages, formulas, and long structured documents. The paper describes DFlash as 'significantly reducing the latency of long structured outputs' while preserving the output distribution, using a lightweight block-diffusion draft model that proposes candidate tokens in parallel for the target model to verify in a single pass. If those numbers hold in production, that reshapes what latency-sensitive OCR pipelines can afford on a lightweight model.

The other piece worth watching is Agentic Data Flow, an agent-driven pipeline that the paper describes as transforming 'model weaknesses into executable data requirements' and then autonomously performing material search, quality verification, and pipeline development. The bet is that once you can get models to identify their own failure modes and construct targeted training sets around them, you can push a small OCR VLM into long-tail territory (ancient scripts, fine-grained charts, multilingual text, hallucination checks) without training a much larger model.

The honest caveat is that this is a paper release, not an independent audit. The paper reports 'top-tier' placement on OmniDocBench v1.6 but the abstract does not give component scores, model size for the 1.5 version, or head-to-head numbers versus specific competitors, so take the specifics as reported, not settled. It also does not say how much of the speedup survives on real-world documents versus the benchmark mix.

If the DFlash gains generalize, the practical upside is aimed squarely at teams shipping OCR at scale: document processors, enterprise back-offices, and anyone running a lightweight VLM on constrained hardware. That is where a 2x-plus vLLM speedup, if it holds, is felt in dollars per page.