We introduce OCRia-1, a 4B vision-language model (VLM) for document parsing that excels at preserving table structure, extracting chart data and retaining text formatting. OCRia-1 handles content that has traditionally been challenging in document parsing: merged cells and spanning headers, tables with thousands of cells, chart data, and meaningful text formatting such as bold, italics, subscripts and superscripts, preserving details needed for business-critical document analysis. At the time of this report, it achieves the highest scores on tables, charts and text formatting among the systems evaluated, including a 35B model and Mistral OCR 4.1. Its mean ParseBench score over the four dimensions we evaluate is 0.87, against 0.75 for the next-best system. We emphasize data quality and reward design in training OCRia-1. Synthetic pages provide exact ground truth for long tables and rich formatting, while real pages are transcribed using a pipeline grounded in the PDF text layer. We attribute OCRia-1's performance primarily to careful data curation that prioritizes quality and diversity over quantity. Dimension-specific rewards evaluate table reconstruction, chart-data extraction and text-formatting fidelity. We train the model on the curated data using supervised fine-tuning (SFT), then refine it through reinforcement learning with verifiable rewards (RLVR). To our knowledge, OCRia-1 is the fastest published document-parsing VLM in its class, decoding at about 4,700 tokens per second with a single request on one B200 GPU, about nine times faster than the base model on stock vLLM. This is achieved through a block-diffusion speculator, quantization and custom kernels.
Figures & tables
Figure 1: How the training set is built. For real pages, we generate transcriptions, check them against the page image and available PDF text, and correct errors before accepting them as ground truth. For synthetic pages, we generate the content first and render it using styles from real documents, so the correct transcription is already known.
Source
Pages
Document class
Internet Archive
23,722
public-domain scans (1800-1975 books on science and mathematics, statistical annexes, encyclopaedias, and letters) and magazines
arXiv
16,142
scientific formula-dense pages
US / UK government statistical publications
12,966
Federal Reserve, IRS SOI, Census, FDIC, BEA, EPA, USDA, EIA.
Contracts and public documents with tracked changes
6,903
CUAD contracts, UN and EMA public documents
Library of Congress
5,199
handwritten manuscripts and early printed matter with volunteer transcriptions
Table 1: Real pages collected by source. We exclude 994 of these pages from training because they exceed the context limit.
Element type
Pages
Tables and charts
13,953
Forms and formatting
3,189
Text variants
894
Table 2: We focus synthetic generation on tables and charts to obtain complex examples with exact ground truth.
Figure 2: Example training pages. Top: real pages from the chart, arXiv, Internet Archive and government statistical collections. Bottom: synthetic pages from the spanned-table, very-long-table, tracked-changes and form generators.
Model
ParseBench ∗
olmOCR-bench †
PulseBench-Tab
fr-bench-pdf2md
Infinity-Parser2-Pro (35B)
0.74
90.2
0.82
54.9
LightOnOCR-3-4B
0.75
90.1
0.82
71.4
olmOCR-2 (7B)
0.43
83.1
0.80
56.9
Chandra-OCR-2 (4B)
0.75
89.4
0.80
65.5
dots.mocr (3B)
0.56
87.4
0.82
56.8
Surya OCR 2 (0.7B)
0.58
85.4
0.71
50.9
Table 3: OCRia-1 leads on four well-known benchmarks, especially in tables, charts and text formatting.
Model
Mean ∗
Tables
Charts
Text formatting
Text content
Infinity-Parser2-Pro (35B)
0.74
0.86
0.61
0.59
0.90
LightOnOCR-3-4B
0.75
0.83
0.66
0.64
0.89
olmOCR-2 (7B)
0.43
0.73
0.14
0.02
0.84
Chandra-OCR-2 (4B)
0.75
0.89
0.65
0.61
0.84
dots.mocr (3B)
0.56
0.85
0.00
0.47
0.90
Surya OCR 2 (0.7B)
0.58
0.80
0.01
0.68
0.84
Table 4: On ParseBench, OCRia-1 leads on tables, charts and text formatting, and ties for best on text content.
Figure 3: OCRia-1 leads in page throughput (left) and decoding speed per request (right) at one (top) and eight (bottom) concurrent requests. Each system runs on one B200. Table 21 gives the exact values.
Appendix figures & tables26 assets
Supplementary material from the paper’s appendix.
Appendix
Group
Rules
What it catches
TABLE RULES 25 rules, asked of every table
Cells
T01–T04
one wrapped cell split into several rows; separately printed records merged into one multi-line cell; invented spacer rows; real blank cells removed or filled
Spans
T05, T07–T09, T23
a category or subtotal label spanned across cells that hold values; wrong grid or span dimensions; repeated values merged into one cell; a real merged cell that loses what it covers; empty regions the rules treat as equivalent flagged as errors
Headers
T10–T12
header tiers flattened, dropped or made up; a header span over the wrong columns; missing or invented headers and columns
Content
T06, T13–T16
sparse or subsection values left out; a cell’s text or number corrupted; superscripts, footnote marks or symbols detached from their value; a value in the wrong row or column; a long table that stops early or repeats
Boundaries
T17–T22
a borderless grid flattened into prose; one table split at a visual band; neighbouring tables woven together; a continuation that invents row relationships; printed occurrences duplicated or suppressed; captions, notes and totals attached to the wrong table
Appendix
Table 5: Rules of the audit. The table rules apply to every table of a draft and the page rules once to the page. The last group is added for pages in non-Latin scripts.
Setting
Value
Rollouts per page
12
Sampling
temperature 0.8, top- p 0.9
Batch
64 pages, 12 rollouts each (768 sequences)
Optimiser
AdamW, learning rate 1×10−6
Gradient clipping
0.5
KL penalty
in the loss, coefficient 0.04
Appendix
Table 6: RLVR settings.
Pages
Share
DOCUMENT TYPE
Synthetic pages from our generators
3,788
33.7%
Curated table, chart and figure collections
2,609
23.2%
Government and regulatory filings (FEC, IRS)
1,799
16.0%
Public document-layout datasets (DocLayNet, RVL-CDIP)
952
8.5%
Forms
803
7.1%
Appendix
Table 7: The 11,234 pages of the training split, by what kind of document each page comes from and by how its reference transcription was made. Document types are inferred from the page identifier, so the groups are reliable and the individual counts approximate. Of the references, 6.4% come from a page’s own text layer and 32.7% from the generator that drew the page; the rest were written by a model.
Implied resolution
Share
∼ 200 dpi
56.2%
∼ 150 dpi
28.7%
∼ 100 dpi
7.4%
Square (non-page aspect)
6.8%
Other ( ≤ 75 dpi)
0.8%
Other ( ≥ 225 dpi)
0.3%
Appendix
Table 8: Page-image resolution in RLVR. The implied resolution is the long edge of the image divided by 11 inches.
Score
1,024 px
1,536 px
PARSEBENCH
Mean
0.83
0.83
Table
0.87
0.88
Chart
0.79
0.79
Text content
0.85
0.85
Text formatting
0.80
0.80
Appendix
Table 9: Training at 1,024 against 1,536 px on the same 30k pages. Qwen3.5-4B, plain transcription, two epochs at batch 64 and learning rate 1×10−5 ; every benchmark served at its usual rendering.
Component
What it scores
Mirrors
WHOLE-PAGE TRANSCRIPTION panels \li@chip bench, \li@chip text
\li@chip bench
benchmark pass rate on the completion, weighted by content type; presence-based, so it rewards recall at no precision cost
olmOCR-bench
\li@chip cct
transcription precision, 1−lev(p,g)/max(∣p∣,∣g∣) on de-tabbed plaintext; penalises extra, reordered and marked-up text
lead table term: mean of detection, cell-index accuracy and cell-content accuracy
SCORE \li@chip score_tables, by direct call
Appendix
Table 10: Reward components that score what the page says: the transcription itself, its tables and charts, and its inline formatting. Panel names are given with each group.
Component
What it scores
Mirrors
REGIONS AND BOXES panels \li@chip layout, \li@chip coverage, \li@chip binding, \li@chip emission, \li@chip bbox
\li@chip layout
F1 of expected against emitted region identifiers
layout element rule
\li@chip coverage
recall of reference regions the output attributed, counting a region as attributed when its marker carries non-empty text; separates attributing a region from transcribing it well
layout element rule
\li@chip binding
of the regions attributed, the fraction whose text is routed to the right identifier, by requiring that its bound text best-match that region’s reference over all regions; targets right content bound to the wrong region
layout attribution
\li@chip grounding_bind
mean soft-distance credit over emitted point markers
\li@chip emission_penalty
fraction of emitted region markers that are waste: an empty body on a text-bearing region, a repeated identifier, or one absent from the region table
Appendix
Table 11: Reward components that score where content is placed and that guard against degenerate output. Panel names are given with each group.
System
What the reward scores
Granularity
Penalties
PUBLISHED MODELS as their authors describe them
Infinity-Parser [ 41 ]
edit distance, paragraph count, reading order
whole page
none described
olmOCR 2 [ 31 ]
binary unit tests, on synthetic pages whose source markup gives the answer; an output that ends with the end-of-sequence token; document metadata
per test, binary
none described; the end-of-sequence reward guards against outputs that never stop
LightOnOCR [ 40 ]
the olmOCR 2 unit tests and formulas that render; in a separate variant, the overlap of the emitted image boxes with the true boxes
per test; per box
repetition loops, malformed formulas, missing and invented boxes
Infinity-Parser2 [ 13 ]
eight tasks jointly, each with its own metric: tree edit distance for cropped tables, value-and-label matching for charts converted to tables, edit distance for full pages
per task
none described
THIS WORK
Appendix
Table 12: The reward of each reinforcement-trained document-parsing VLM, as its authors describe it. “None described” records that the published account does not mention penalties against degenerate output, not that the system has none.
Stage
Measurement
Result
Rendering
image tokens per table cell, and type height
63 tokens per cell and 20 px type on pages up to 100 cells; 1.1 tokens per cell and 8 px above 1,000 cells
Vision encoder
probe accuracy for character identity as the page gets fewer image tokens
0.82 at 3,796 tokens, 0.40 at 1,024, 0.22 at 400, 0.11 at 144 (chance floor about 0.09)
Projector
probe accuracy before and after the projector
digits 0.84 → 0.76, letters 0.56 → 0.41; gutters, column boundaries and merged cells drop by at most 0.02
Decoder
activation patching on degraded pages
clean image tokens at the decoder input restore all of the lost quality; at layer 16 they restore 50% (95% CI 39–63%)
Appendix
Table 13: Where small text is lost, stage by stage, on the Qwen3.5-4B model.
Cells
Pages
Table score
Image tokens per cell
Type height (px)
Reference length (tokens)
Abnormal end
1–100
259
0.96
63.0
19.6
596
0
101–300
142
0.92
12.6
18.6
2,368
0
301–600
61
0.82
5.0
14.6
5,111
1
601–1,000
28
0.88
3.2
13.2
9,196
0
Over 1,000
13
0.59
1.1
8.0
24,139
6
Appendix
Table 14: ParseBench table pages by size (4B model, 150 dpi, output limit 24,576 tokens). The table score is the mean over pages; the other columns are medians.
Image tokens
Page image
olmOCR-bench
Long tiny text
266
1,288 px render (shrunk)
45.7
7.9
532
1,288 px render (shrunk)
67.2
36.9
1,240
1,288 px render
83.5
92.3
2,268
1,288 px render (enlarged)
84.4
93.9
2,451
1,822 px render
85.9
94.2
Appendix
Table 15: olmOCR-bench for the 4B model as the image-token budget changes. Image tokens are the median per page.
Model
arXiv math
Old scans math
Tables
Old scans
Headers, footers
Multi-column
Long tiny text
Base
Overall
Infinity-Parser2-Pro
88.1
91.3
91.2
58.2
95.8
83.7
92.5
99.9
87.6
LightOnOCR-3-4B
91.1
89.1
90.2
52.5
89.2
85.6
93.2
99.5
86.3
olmOCR-2 (7B)
82.9
82.1
84.3
48.3
95.7
84.3
81.4
99.7
82.3
Chandra-OCR-2 (4B)
86.9
89.1
92.1
51.1
91.4
82.1
93.7
99.9
85.8
dots.mocr (3B)
85.9
85.5
90.7
48.2
94.0
85.3
81.6
99.7
83.9
Surya OCR 2 (0.7B)
88.3
81.4
86.6
41.8
92.5
82.4
93.7
99.7
83.3
Appendix
Table 16: olmOCR-bench by category. Overall is the official average over the eight columns. The other systems’ rows are the numbers they publish (the LightOnOCR-3 post; olmOCR-2’s model card); the OCRia-1 row is our run. LightOnOCR-2-1B’s paper reports 83.2 with the headers-and-footers category left out; here its overall is recomputed over all eight categories, using the headers-and-footers score (19.7) from the same paper.
Model
arXiv math
Old scans math
Tables
Old scans
Headers, footers
Multi-column
Long tiny text
Base
Overall
Infinity-Parser2-Pro
82.9
85.2
87.7
51.0
92.1
82.8
84.4
99.9
83.2
LightOnOCR-3-4B
86.9
78.8
87.6
48.7
87.6
84.3
83.5
99.7
82.1
olmOCR-2 (7B)
82.9
82.1
84.3
48.3
95.7
84.3
81.4
99.7
82.3
Chandra-OCR-2 (4B)
80.3
84.3
86.7
45.8
90.8
80.7
81.2
99.9
81.2
dots.mocr (3B)
77.9
78.4
86.3
40.9
90.0
82.2
85.3
99.6
80.1
Surya OCR 2 (0.7B)
78.8
76.0
82.1
38.2
93.2
79.9
73.3
99.7
77.6
Appendix
Table 17: olmOCR-bench: our runs at 1,288 px alongside published baselines. For our runs, pages are rendered at 1,288 px on the long side, models are served with vLLM at their own image budgets, and outputs receive the same post-processing and are scored by the olmOCR-bench scorer. Mistral OCR 4.1’s row is from the LightOnOCR-3 post and olmOCR-2’s from its model card at the same 1,288 px. Overall is the official average over the eight columns. Of the 440 arXiv-math tests that OCRia-1 fails, 143 render to the same formula as the reference but write the L a T e X in a form other than the KaTeX one the test expects; counting them as passed gives 89.9 on that category.
Model
Mean
Coverage %
Median
Perfect tables
Infinity-Parser2-Pro
0.82
96.8
0.93
419 (23.0%)
LightOnOCR-3-4B
0.82
96.6
0.93
431 (23.7%)
olmOCR-2 (7B)
0.80
92.4
0.90
354 (19.5%)
Chandra-OCR-2 (4B)
0.80
93.0
0.92
340 (18.7%)
dots.mocr (3B)
0.82
88.7
0.92
308 (16.9%)
Surya OCR 2 (0.7B)
0.71
93.7
0.82
224 (12.3%)
Appendix
Table 18: Results on PulseBench-Tab, 1,820 tables. Mean and median T-LAG score over the tables scored; coverage is the share of tables with a parseable prediction.
fr-bench-pdf2md, pass rate by category
Model
Forms
Graphics
Handwritten
Long table
Long table 2
Multi-column
Tiny text
Overall
Infinity-Parser2-Pro
33.9
26.9
33.3
51.5
50.0
76.4
78.7
54.9
LightOnOCR-3-4B
49.6
68.3
41.2
80.6
70.7
83.6
83.2
71.4
olmOCR-2 (7B)
43.8
37.7
20.6
67.9
63.9
84.8
66.2
56.9
Chandra-OCR-2 (4B)
38.3
72.8
22.4
80.6
77.8
70.9
70.0
65.5
dots.mocr (3B)
40.1
26.1
13.9
56.7
59.8
80.6
79.3
56.8
Appendix
Table 19: Results on fr-bench-pdf2md, pass rate by category and overall, over 2,318 tests.
Concurrent requests
Output tok/s per request
Aggregate output tok/s
Pages/s
1
4,734
–
2.01
8
1,932
9,964
7.42
16
1,103
12,740
8.67
32
566
14,833
9.77
48
400
15,240
9.98
64
302
15,557
10.21
Appendix
Table 20: Speed of OCRia-1 on one B200 with the final serving stack of Appendix E , against the number of concurrent requests. Measured with AIPerf, each value is the mean of three runs. Aggregate output tokens per second is all output tokens of a run over the run’s wall-clock time. Section E.1 gives the protocol.
Figure 4: Single-stream decoding speed on one B200. The same model in bf16 on stock vLLM against the optimized stack (NVFP4 target, DFlash2 draft with 31 draft tokens, the kernels of Appendix E ), greedy decoding, the production prompt, one page at a time. Bars are the end-to-end rate, time to first token included, of the median request at a client in the same region; the whisker is the range seen from a browser client over runs; the diamond is the AIPerf per-request decode rate on the concurrency ladder, which excludes the time to first token.
Model
Concurrent requests
Output tok/s per request
Pages/s
OCRia-1 (optimized)
1
4,734
2.01
Nemotron-3.5-Parse
1
1,395
0.80
HunyuanOCR-1.5 + DFlash
1
1,403
0.71
SmolDocling-256M
1
894
0.55
Falcon-OCR bf16
1
599
0.39
OCRia-1 (optimized)
8
1,932
7.83
Appendix
Table 21: Cross-model speed comparison on one B200 with one and eight concurrent requests. Table 20 reports the separate concurrency sweep.
Setting
k
Tokens per step
Accepted per step
Condition
OCRIA-1, DFLASH 2 DRAFT, GREEDY tokens emitted per verification step; accepted drafts of 31
Tuning pages, c = 8
31
17.5
16.5
timed inside the server, B200
Production pages, 40-page sample, c = 1
31
12.5
11.5
timed inside the server, B200
Production traffic, 512 pages, c = 8
31
13.3
–
per-request mean, AIPerf ladder of Table 20
Fixed canary pages at container start, c = 1
31
14.2
–
B200
The same canary pages on a B300
31
12.9–13.2
–
B300 arithmetic differs from B200; 14.2 on B200
Appendix
Table 22: Accepted tokens per verification step. Top group: the served DFlash 2 draft (6 layers, block 32, selector top-16) on the OCRia-1 NVFP4 target at k = 31, temperature 0; a step emits the accepted drafts plus one target token, so 17.5 tokens per step means 16.5 of 31 drafts accepted. Tuning pages are the eight pages the kernel work was tuned on; production pages are a 40-page production sample and the 512-page AIPerf ladder of Table 20 ; the canary is a fixed page set run at every container start. Bottom group: published figures, each with its authors’ definition of accepted length.
Figure 5: Acceptance by draft position. Probability that the target accepts at least the first p draft tokens, for the greedy walk over the draft’s candidates (the served dynamic-programming walk accepts slightly more) and for an oracle that picks the right token whenever it is among the draft’s sixteen candidates, with the hit rate at each position given the prefix was right; offline replay of 9,193 greedy draft calls on 100 pages with 31 draft tokens. The area under the walk curve is the accepted tokens per step.
Figure 6: A dense financial table. A page of a Texas insurance rate filing, with a three-level header: a title row over the years 2011 to 2013, and an amount and a percent column under each year. OCRia-1 writes the page as one HTML table, with the header spans as \li@chip colspan attributes and the section rows ( COUNTRYWIDE ) kept in place.
Figure 7: A chart transcribed as its data. A line chart with three series over five years. OCRia-1 writes the chart as a table of its 15 data points, one row per year and one column per series, and the values match the labels printed on the chart.
Figure 8: Text formatting. A hearing notice that marks a withdrawn and a continued case by striking them through, underlines the case names and sets the times in bold. OCRia-1 keeps all three: the struck-through entries as Markdown strikethrough, the case names as \li@chip <u> and the times as bold.
Figure 9: Handwritten formulas. A handwritten note in Czech on the derivative of a composite function. OCRia-1 transcribes the Czech text and writes each formula as L a T e X .
Figure 10: A table longer than the output budget. A page that holds one table of vehicle codes, 195 rows and 18 columns, about 3,500 cells. OCRia-1 keeps the 18 columns and writes the rows in order, without repeating any, until it reaches the 24,576-token output limit after 183 rows. The last 12 rows are missing and the table has no closing tags. The left panel shows the bottom of the page and the right panel the end of the output.
Figure 11: A chart without printed values. A bar chart of monthly listings for two ransomware groups, with no value printed on the bars. OCRia-1 transcribes it as a table of all 24 bars, reading each value from the axis. On six of the ten points that ParseBench tests, the reading is one or two units from the expected value, for example 58 against 60 and 98 against 100 for LockBit in January and February, and the benchmark counts each of these as a miss.
We present Qianfan-OCR, a 4B-parameter end-to-end vision-language model that unifies document parsing, layout analysis, and document understanding within a single architecture. It performs direct image-to-Markdown conversion and supports diverse prompt-driven tasks including table extraction, chart understanding, document QA, and key information extraction. To address the loss of explicit layout analysis in end-to-end OCR, we propose Layout-as-Thought, an optional thinking phase triggered by special think tokens that generates structured layout representations -- bounding boxes, element types, and reading order -- before producing final outputs, recovering layout grounding capabilities while improving accuracy on complex layouts. Qianfan-OCR ranks first among end-to-end models on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), achieves competitive results on OCRBench, CCOCR, DocVQA, and ChartQA against general VLMs of comparable scale, and attains the highest average score on public key information extraction benchmarks, surpassing Gemini-3.1-Pro, Seed-2.0, and Qwen3-VL-235B. The model is publicly accessible via the Baidu AI Cloud Qianfan platform.
We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. We build a data engine that combines filtered real-document annotations with synthetic pages whose rendered images and Markdown targets are derived from the same HTML source. The training recipe includes supervised fine-tuning, reinforcement learning on a 4B branch with a multi-component reward design, on-policy distillation into the 0.8B model, and model fusion. On OmniDocBench v1.6, OvisOCR2 achieves a state-of-the-art overall score of 96.58, placing an end-to-end model at the top of this leaderboard previously dominated by pipeline methods and highlighting the potential of end-to-end document parsing. On PureDocBench, OvisOCR2 also achieves the highest Avg3 score of 75.06. Beyond these two public benchmarks, we evaluate OvisOCR2 on an in-house benchmark designed to cover a broader set of long-tail and challenging scenarios. OvisOCR2 obtains the best overall performance among the compared methods, providing further evidence of its generalization and robustness. OvisOCR2 is available at https://huggingface.co/ATH-MaaS/OvisOCR2.
Compact OCR-specific vision-language models achieve strong document parsing performance, but often rely on costly supervision and focus primarily on visual-text reconstruction. We introduce Xiaomi-OCR-0, a unified 0.8B model for document parsing and OCR-centric understanding. We build an approximately 170M-sample OCR-centric corpus using an automated data engine that combines expert consensus, render-based verification, and targeted synthesis. Starting from Qwen3.5-0.8B, our progressive training recipe combines Q-Mask-based text anchoring, continued pretraining, and mixed-task reinforcement learning (Mix-RL). Xiaomi-OCR-0 achieves 95.24 on Real5-OmniDocBench, 96.83 on OmniDocBench v1.6, and 87.94 on Wild-OmniDocBench, while reaching an average score of 83.2 across five OCR-oriented VQA benchmarks. Ablations further show that, with sufficient parsing training, OCR-centric understanding supervision provides additional gains for document parsing. Homepage: https://huggingface.co/spaces/SeerRay-Lab/Xiaomi-OCR-0.