Source-linked AI summary
Jina-OCR-v1: Efficient Document Parsing with Speculative Decoding and Dense Verifiable Rewards
Alejandro Barón García, Feng Wang, Emilia Garcia Casademont, Han Xiao
TL;DR
End-to-end document parsing can unify layout, text, formula, and table recovery, but decoding cost and post-training data coverage constrain production use, especially on low-budget GPUs. Jina-OCR-v1 combines a compressed-vision, 3B-MoE backbone with shared-block FastMTP and dense verifiable-reward training. It reports benchmark scores of 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and FastMTP nearly doubles decoding speed on an NVIDIA L4.
Problem
Production use of end-to-end document parsing, which unifies layout, text, formula, and table recovery, remains constrained by decoding cost and post-training data coverage, especially on low-budget GPUs.
Source: §1 · p. 1Method
Jina-OCR-v1 combines a compressed-vision, 3B-MoE backbone with recursively shared FastMTP drafting and post-training under dense verifiable rewards.
Source: §1 · p. 2Results
Jina-OCR-v1 scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, is the fastest system measured in pages per second, and FastMTP nearly doubles decoding speed on an NVIDIA L4.
Source: §7 · pp. 11–12Takeaways & Limitations
Page throughput and token throughput rank systems differently because systems emit different-length transcriptions, and output length can be optimized separately from parsing quality.
Source: §7 · p. 11
Abstract
from arXiv · showhide
We present Jina-OCR-v1, an end-to-end document parsing model built to serve on low-budget GPUs. It combines the compressed-vision encoder and the 3B mixture-of-experts decoder of DeepSeek-OCR, which activates about 570M parameters per token, with a FastMTP speculative decoding head that shares a single draft block recursively across K=3 prediction steps. Greedy verification makes decoding lossless. Post-training combines instruction alignment, robustness fine-tuning on difficult documents, and GRPO under dense verifiable rewards: deterministic formula, table, and structural checks that award partial credit. The training data mixes cleaned public corpora with targeted synthetic pages. At the default dynamic-resolution setting, Jina-OCR-v1 scores 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench, and reaches the highest page throughput in our comparison at 2.57 pages per second. On a low-budget GPU such as the NVIDIA L4, FastMTP doubles decoding speed over greedy autoregressive decoding. The model is publicly available at https://huggingface.co/jinaai/jina-ocr-v1.
1 Introduction
Jina-OCR-v1 combines compressed vision, speculative decoding, and dense verifiable rewards to address decoding cost and post-training coverage in end-to-end document parsing. It lies on the accuracy frontier against its active-parameter budget on both benchmarks and is the most compact point on the OmniDocBench frontier.
1 Introduction
- Production use, especially on low-budget GPUs, is constrained by decoding cost and post-training data coverage, while end-to-end parsing combines layout, text, formula, and table recovery.Open labels may contain degeneration loops or malformed structures, and formula- and table-specific rewards apply only to pages containing those elements.Source: §1 · p. 1
- With 570M decoder parameters active per token, Jina-OCR-v1 lies on the accuracy frontier against that budget on both benchmarks and is the most compact point on the OmniDocBench frontier.Figure 1 compares benchmark overall against active parameters.Source: Figure 1 · p. 2, §1 · pp. 2–3
- FastMTP recursively shares one draft block across K=3 prediction steps on the compressed-vision, 3B-MoE DeepSeek-OCR backbone, keeping draft parameter count constant in K.The target is the remaining autoregressive bottleneck, exploiting local structure in OCR output to draft several tokens before verification.Source: §1 · pp. 1–2
- GRPO uses multiplicative, deterministic reference-based rewards for content, structure, unit tests, repetition, and format, giving partially correct pages partial credit.Synthetic JinaOCRSynth pages packed with scorable formulas and tables increase the share of training pages where formula and table reward terms apply.Source: §1 · p. 2
- Instruction-oriented training covers multiple parsing styles, element transcription, captioning, document VQA, and key-information extraction, with generated instructions and auxiliary grounding examples.The mixture builds on ReaderLM-v2.Source: §1 · p. 2
- The inherited DeepEncoder maps 4,096 patches from a 1024×1024 global view to 256 tokens, so comparable-resolution pages use far fewer visual tokens.A patch side of p=64 represents about 4,096 pixels per token; dynamic-resolution Gundam mode is not specified in this passage.Source: §1 · p. 2
2 Related Work
Related work spans end-to-end document parsers, efficient multi-token decoding, and reinforcement learning with verifiable rewards. Jina-OCR-v1 builds on shared-block drafting and combines multiplicative rewards with per-term floors, repetition penalties, and synthetic formula and table pages.
End-to-end document parsing
- End-to-end document parsing ranges from academic OCR and broad OCR task suites to recent end-to-end systems and two-stage pipelines that first run layout analysis.Nougat [Blecher et al., 2023] demonstrated end-to-end academic document OCR; GOT-OCR2.0 [Wei et al., 2024] broadened the task suite.Source: §2 · p. 3
Efficient decoding and multi-token prediction
- Multi-token prediction methods include parallel or sequential draft heads, headless lookahead, feature-level drafting, and shared-block FastMTP [Cai et al., 2025].The paper follows the shared-block line for constant draft parameter count and trains it with feature- and distribution-alignment losses.Source: §2 · p. 3
Reinforcement learning for document parsing
- Document parsing RL uses deterministic signals such as edit distance, TEDS [Zhong et al., 2020], and CDM [Wang et al., 2024a]; this work adds reward floors, a ReMax baseline [Li et al., 2023], repetition penalties, and synthetic formula/table pages.The passage notes that structured document elements permit deterministic scoring from reference transcriptions without a reward model or judge.Source: §2 · p. 3
3 Model Architecture
Jina-OCR-v1 retains DeepSeek-OCR’s compressed-vision, 3B-MoE backbone and adds a recursively shared FastMTP draft head. Its training alignment and greedy verification let the head draft multiple tokens while preserving the verifier’s greedy output.
3 Model Architecture
- Jina-OCR-v1 retains DeepSeek-OCR’s compressed-vision, 3B-MoE backbone and adds FastMTP, with approximately 570M decoder parameters active per token.At 1024×1024, DeepEncoder compresses 4,096 global patches to 256 tokens; dynamic Gundam mode adds at most nine 100-token local tiles.Source: §3 · p. 3, Table 1 · p. 4, Figure 2 · p. 4
3.1 Draft-head architecture
- One shared FastMTP draft block recursively proposes K=3 tokens, keeping draft parameter count constant in K.The head is added to the compressed-vision DeepSeek-OCR backbone.Source: §3 · p. 3, §3.1 · p. 4, Figure 2 · p. 4
- Training aligns recursive depths so each row targets the same verifier state, while an inference-time diagonal advances the target by one token per step.Token embeddings, output norm, and LM head are tied to verifier counterparts; prompt rows retain base states across depths, matching the committed-token KV cache.Source: §3.1 · pp. 4–5, Figure 3 · p. 5
- Greedy verification accepts the longest matching draft prefix, replaces the first rejected draft with the verifier token, and preserves the verifier’s greedy sequence exactly.No stochastic acceptance or residual-distribution resampling is used.Source: §3.1 · p. 5
4 Training Data
The training data combines cleaned public document corpora with targeted synthetic pages and task-specific instructions. Collection and synthesis address difficult examples and increase coverage of structures used by the reward functions.
4 Training Data
- The training mixture combines cleaned public corpora with targeted synthetic pages, pairs samples with task instructions, and uses model error analysis to guide collection.The data section frames the collection around label quality and availability of structures covered by reward functions.Source: §4 · p. 5, §4.1 · p. 5
4.1 Open data collection and cleaning
- Public data spans document OCR and specialized corpora, while historical newspapers, archives, pension files, periodicals, and Japanese wild-text samples add difficult real-world pages.Cleaning removes detected degeneration loops, exact duplicate images, and samples shorter than 64 tokens; selected sources are relabeled or audited with vision-language models.Source: §4.1 · pp. 5–6
- Classifier-driven retrieval targets pages associated with failure modes using handcrafted features, model embeddings, or both, without scoring every candidate with the current model.The classifiers are modality-specific and can flag cues such as table-structure indicators and distinctive LATEX glyphs.Source: §4.1 · p. 6
4.2 Targeted synthesis
- Synthetic pages add layout diversity and difficult examples, while formula- and table-rich pages make structural GRPO rewards applicable more often.Their unit tests cover formula and table structures, including spanned headers, dense numeric cells, and transposed or borderless layouts.Source: §4.2 · p. 6
4.3 Instruction coverage
- Instructions cover full-page parsing, table and formula transcription, captioning, document VQA, and key-information extraction, with English and Chinese variants.Grounding and VQA provide auxiliary visual coverage; OCR is the deployment target.Source: §4.3 · p. 6
5 Post-Training
Post-training repeatedly applies alignment SFT, robustness fine-tuning, and GRPO, then fits FastMTP to the selected verifier. The stages address instruction following, difficult-document robustness, verifiable rewards, checkpoint selection, and draft-head training.
5 Post-Training
- Alignment SFT, robustness SFT, and GRPO recur across training rounds, while FastMTP is fitted once on the resulting selected verifier.Table 3 summarizes the four training objectives.Source: §5 · p. 6, Table 3 · p. 7
5.1 Alignment SFT and length extrapolation
- Alignment SFT extends usable response positions by randomly shifting response position IDs with probability 0.4, while sampling across roughly 15 domains with balanced rates.Prompt positions are untouched, IDs are capped at 32,768, and the shift draws from [1, min(5L, Ltrain/4)].Source: §5.1 · p. 7
5.2 Robustness fine-tuning
- Robustness fine-tuning unfreezes the vision tower at 10^-6, disables packing, corrupts historical scans, standardizes equation formatting, and upweights degraded and structure-rich pages.Corruptions include skew, rotations, downscaling, noise, erosion, and dilation; failure modes identified after each round are oversampled.Source: §5.2 · p. 7
5.3 GRPO under dense verifiable rewards
- GRPO uses deterministic, graded reward terms for content, formulas, tables, structure, unit tests, repetition, and format, with floors preventing a single failed check from zeroing the product.Structure-focused runs use DAPO with eight rollouts and a ReMax greedy-reward baseline to avoid amplifying small within-group variance in near-deterministic OCR rollouts.Source: §5.3 · p. 7, Table 4 · p. 8
5.4 Auxiliary regularizers
- Auxiliary regularizers include load balancing, gate entropy, and logit penalties; a local contrastive term encourages separation between nearby token states.The contrastive term follows the regularization used in ReaderLM-v2 [Wang et al., 2025].Source: §5.4 · p. 8
5.5 Agentic checkpoint merging and data curation
- A fixed-budget agent searches checkpoint merges, evaluates them with unit tests and edit distance, then feeds analyzed failure modes into the next training round.Alignment, robustness fine-tuning, and GRPO repeat across outer-loop rounds; FastMTP is fitted after verifier selection.Source: §5.5 · pp. 8–9
5.6 Draft-head objective
- The draft head is fitted to the selected frozen verifier using state-shift alignment, with shared row-aligned token and state targets across recursive depths.Verifier states and distributions are detached; the verifier backbone and tied embedding, output norm, and LM head remain fixed.Source: §5.6 · p. 9
6 Evaluation
Jina-OCR-v1 ranks among the leading specialized systems on both parsing benchmarks and achieves the highest page throughput in the reported comparison. On an NVIDIA L4, FastMTP speeds decoding over the eager autoregressive baseline, while the best speculative depth depends on execution mode.
6.1 Benchmarks and metrics
- olmOCR-Bench scores page-level text, reading order, math, and tables; OmniDocBench v1.6 averages text, formula CDM, and table TEDS.Serving throughput uses a larger model pool on olmOCR-Bench under batched inference.Source: §6.1 · p. 9
6.2 Parsing quality
- 91.14 on OmniDocBench v1.6 and 83.4 on olmOCR-Bench place Jina-OCR-v1 third among specialized models with published overall scores on each benchmark.It is 7.4 points above its DeepSeek-OCR backbone on olmOCR-Bench and ahead of DeepSeek-OCR-2 on OmniDocBench v1.6.Source: §6.2 · p. 9, Table 6 · p. 10, Table 5 · p. 10
6.3 Serving throughput
- Jina-OCR-v1 reaches the highest page throughput by combining competitive token throughput with the shortest outputs among systems scoring above 83.Surya OCR 2 leads token throughput at 3,760 tok/s but emits 3,568 tokens per page; Jina-OCR-v1 emits 1,085 and finishes pages faster.Source: §6.3 · p. 10, Table 7 · p. 11, Figure 5 · p. 11
6.4 Speculative decoding
- On an NVIDIA L4, FastMTP at k=3 reaches 1.95× the eager autoregressive baseline, while CUDA-graph execution makes k=1 best at 1.17×.CUDA graphs raise the autoregressive baseline from 42.7 to 158.3 tok/s; at k=3, acceptance is 2.73 in eager mode and 2.74 in graph mode, while conditional acceptance declines from 0.83 to 0.65 across draft positions.Source: §6.4 · pp. 10–11
7 Conclusion
Jina-OCR-v1 scores strongly across the reported benchmarks and is the fastest measured system by page throughput. On an NVIDIA L4, FastMTP nearly doubles decoding speed; page and token throughput rank systems differently because output lengths differ.
7 Conclusion
- 83.4 on olmOCR-Bench and 91.14 on OmniDocBench v1.6; Jina-OCR-v1 is the fastest measured system in pages per second.The conclusion reports these benchmark scores and the page-throughput comparison.Source: §7 · p. 11
- Because faster-decoding systems emit longer transcriptions, page- and token-throughput rankings differ; output length is independent of parsing quality and can be optimized separately.Source: §7 · p. 11
- Table 8 reports batch-size-1 decoding efficiency on an NVIDIA L4, and its values are not comparable with Table 7.Its speedup baseline is k=0 within the same execution mode.Source: Table 8 · p. 12
- Nearly double decoding speed on an NVIDIA L4, with acceptance behavior invariant to verifier execution mode.The reported speedup is for FastMTP; Table 8 measures batch size 1 on an L4 against the k=0 baseline in the same execution mode.Source: §7 · p. 12, Table 8 · p. 12