BaiduSpecialist modelOpen weights
Unlimited OCR 2
Summary
Unlimited OCR 2 parses multi-page documents with an approach that avoids keeping the full visual history in GPU KV cache. Baidu publishes separate detailed results for MPDocBench, MarathonDocBench, and OmniDocBench v1.6. The dataset card lists about 18,000 pages across 100 MarathonDocBench documents of 100 to 300 pages. The technical report lists 103 to 300 pages for its 100 documents, with a median of 178 pages. The benchmarks cover different tasks and report separate metrics.
Lower is better for edit-distance metrics; the other table metrics are marked higher-is-better. The technical table separates standard decoding from unlimited-horizon decoding, so memory and throughput values apply only to their stated batch configurations. The documented example loads BF16 weights on an A800 with 80 GB; third-party quantizations are not the configurations measured in these tables.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 30 published benchmark results
Page 1 of 3
Measurements without matching peer values
These measurements have no matching peer values under the same test conditions. Their original values and sources remain available here.
MPDocBench Overall (Show measurement, test conditions, and source)
- Source value
- 84.59
- Score
- 84.59
- Metric
- Overall score
- Unit
- %
- Benchmark version
- MPDocBench, technical report Table 1
- Category
- document-parsing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MPDocBench; up to 30-page multi-page document parsing.
- Context
- Publisher-reported result. Do not compare this table directly with a different benchmark suite.
MPDocBench FormulaCDM (Show measurement, test conditions, and source)
- Source value
- 89.39
- Score
- 89.39
- Metric
- FormulaCDM score
- Unit
- %
- Benchmark version
- MPDocBench, technical report Table 1
- Category
- document-parsing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MPDocBench; up to 30-page multi-page document parsing.
- Context
- Publisher-reported result. Do not compare this table directly with a different benchmark suite.
MPDocBench TableTEDS (Show measurement, test conditions, and source)
- Source value
- 87.78
- Score
- 87.78
- Metric
- TableTEDS score
- Unit
- %
- Benchmark version
- MPDocBench, technical report Table 1
- Category
- document-parsing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MPDocBench; up to 30-page multi-page document parsing.
- Context
- Publisher-reported result. Do not compare this table directly with a different benchmark suite.
MPDocBench T-TableTEDS (Show measurement, test conditions, and source)
- Source value
- 89.54
- Score
- 89.54
- Metric
- T-TableTEDS score
- Unit
- %
- Benchmark version
- MPDocBench, technical report Table 1
- Category
- document-parsing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MPDocBench; up to 30-page multi-page document parsing.
- Context
- Publisher-reported result. Do not compare this table directly with a different benchmark suite.
MPDocBench FigureF1 (Show measurement, test conditions, and source)
- Source value
- 72.91
- Score
- 72.91
- Metric
- FigureF1 score
- Unit
- %
- Benchmark version
- MPDocBench, technical report Table 1
- Category
- document-parsing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MPDocBench; up to 30-page multi-page document parsing.
- Context
- Publisher-reported result. Do not compare this table directly with a different benchmark suite.
MPDocBench HeadingTEDS (Show measurement, test conditions, and source)
- Source value
- 60.1
- Score
- 60.1
- Metric
- HeadingTEDS score
- Unit
- %
- Benchmark version
- MPDocBench, technical report Table 1
- Category
- document-parsing
- Direction
- Higher is better
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MPDocBench; up to 30-page multi-page document parsing.
- Context
- Publisher-reported result. Do not compare this table directly with a different benchmark suite.
MarathonDocBench Overall (Show measurement, test conditions, and source)
- Source value
- 77.06
- Score
- 77.06
- Metric
- Overall score
- Unit
- %
- Benchmark version
- MarathonDocBench, technical report Table 2
- Category
- document-parsing
- Direction
- Higher is better
- Sample
- 100
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MarathonDocBench; 100 documents spanning 103 to 300 pages (median 178 pages), with 205,000 annotated layout elements.
- Context
- Sample count is 100 benchmark documents, not model runs. Edit-distance metrics are lower-is-better.
MarathonDocBench FormulaCDM (Show measurement, test conditions, and source)
- Source value
- 87.93
- Score
- 87.93
- Metric
- FormulaCDM score
- Unit
- %
- Benchmark version
- MarathonDocBench, technical report Table 2
- Category
- document-parsing
- Direction
- Higher is better
- Sample
- 100
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MarathonDocBench; 100 documents spanning 103 to 300 pages (median 178 pages), with 205,000 annotated layout elements.
- Context
- Sample count is 100 benchmark documents, not model runs. Edit-distance metrics are lower-is-better.
MarathonDocBench TableTEDS (Show measurement, test conditions, and source)
- Source value
- 74.81
- Score
- 74.81
- Metric
- TableTEDS score
- Unit
- %
- Benchmark version
- MarathonDocBench, technical report Table 2
- Category
- document-parsing
- Direction
- Higher is better
- Sample
- 100
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MarathonDocBench; 100 documents spanning 103 to 300 pages (median 178 pages), with 205,000 annotated layout elements.
- Context
- Sample count is 100 benchmark documents, not model runs. Edit-distance metrics are lower-is-better.
MarathonDocBench T-TableTEDS (Show measurement, test conditions, and source)
- Source value
- 71.11
- Score
- 71.11
- Metric
- T-TableTEDS score
- Unit
- %
- Benchmark version
- MarathonDocBench, technical report Table 2
- Category
- document-parsing
- Direction
- Higher is better
- Sample
- 100
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MarathonDocBench; 100 documents spanning 103 to 300 pages (median 178 pages), with 205,000 annotated layout elements.
- Context
- Sample count is 100 benchmark documents, not model runs. Edit-distance metrics are lower-is-better.
MarathonDocBench FigureF1 (Show measurement, test conditions, and source)
- Source value
- 71.79
- Score
- 71.79
- Metric
- FigureF1 score
- Unit
- %
- Benchmark version
- MarathonDocBench, technical report Table 2
- Category
- document-parsing
- Direction
- Higher is better
- Sample
- 100
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MarathonDocBench; 100 documents spanning 103 to 300 pages (median 178 pages), with 205,000 annotated layout elements.
- Context
- Sample count is 100 benchmark documents, not model runs. Edit-distance metrics are lower-is-better.
MarathonDocBench HeadingTEDS (Show measurement, test conditions, and source)
- Source value
- 48.2
- Score
- 48.2
- Metric
- HeadingTEDS score
- Unit
- %
- Benchmark version
- MarathonDocBench, technical report Table 2
- Category
- document-parsing
- Direction
- Higher is better
- Sample
- 100
- Source type
- vendor-reported
- Evaluator
- Baidu Unlimited OCR 2 technical report, benchmark tables
- Status
- active
- Retrieved at
- 2026-10-29
- Methodology
- Vendor-reported result. End-to-end VLM result for Unlimited OCR 2 on MarathonDocBench; 100 documents spanning 103 to 300 pages (median 178 pages), with 205,000 annotated layout elements.
- Context
- Sample count is 100 benchmark documents, not model runs. Edit-distance metrics are lower-is-better.
Page 1 of 3
All specifications
Architecture
Licence and access
Further details
- Attention mechanism
- Sparse Reference Sliding Window Attention keeps a bounded set of reference tokens active; the model card gives 512 tokens as an example.
- Source: Attention mechanism
- Multi-page documents
- A document is passed as an ordered sequence of page images in one parsing request.
- Source: Multi-page documents
- Documented inference hardware
- NVIDIA A800 with 80 GB and BF16 weights in the Transformers example
- Source: Documented inference hardware
- MarathonDocBench scope in the report
- Technical report: 100 documents of 103 to 300 pages (median: 178 pages) and 205,000 annotated layout elements
- Source: MarathonDocBench scope in the report
- MarathonDocBench scope in the dataset card
- Dataset card: 100 documents of 100 to 300 pages each and about 18,000 pages in total
- Source: MarathonDocBench scope in the dataset card
- Known limitation
- The technical report notes limited training-data diversity and limited global document understanding, for example of heading hierarchies.
- Source: Known limitation
- MarathonDocBench documents
- 100 documents; about 18,000 pages in total according to the dataset card
- Source: MarathonDocBench documents
Sources and data date
Benchmark sources at a glance
Which independent evaluators tested this model. This is coverage, not a quality ranking.
No active result from the reviewed external evaluators can currently be attributed to this model. Other source evidence remains available below.
Without vendor claims and historical or disputed results; those stay labelled in the single values.
| Type | Evidence and data date |
|---|---|
| Research status | Research date October 29, 2026. 60 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | Exact accuracy curve by document length Note Figure 6 plots accuracy across length bins without a numeric data table. Only exact values printed in the benchmark tables are recorded. |
| Not published | Absolute decoding latency at 100 pages Note Figure 6 prints relative speedups at 100 pages but does not tabulate an exact absolute latency. |
| Not found in the inspected sources | Hardware for Table 4 feasibility values Note The report specifies batch conditions for Table 4, but the table does not identify its GPU. The separately documented A800 BF16 setup is not attributed to these measurements. |
| Not found in the inspected sources | Vals rows in captured payloads Note Nine complete Vals payloads dated 28 October were checked; no exact model ID or unambiguous model label. Limited to these saved benchmarks. |
| Unresolved | Uncaptured Arena Code/Speech leaderboards Note Text, Agent including Code/Chat/Work, and twelve WebDev views were captured. A separate general Code and Arena Speech leaderboard were not saved as complete tables. |
| Unresolved | Artificial Analysis catalog reconciliation Note The full current AA model and benchmark catalog was not re-reconciled for this profile in this sidecar. This sidecar makes no AA absence claim; native AA imports remain authoritative. |
| Additional source | Baidu Unlimited OCR 2 model card (retrieved October 29, 2026) · Baidu Unlimited OCR 2 model card · Editorial description reviewed October 29, 2026 |
| Additional source | Baidu Unlimited OCR repository (retrieved October 29, 2026) · Baidu Unlimited OCR repository · Editorial description reviewed October 29, 2026 |
| Additional source | Baidu Unlimited OCR 2 technical report and benchmark tables (retrieved October 29, 2026) · Baidu Unlimited OCR 2 technical report and benchmark tables · Editorial description reviewed October 29, 2026 |
| Additional source | Baidu MarathonDocBench dataset card (retrieved October 29, 2026) · Baidu MarathonDocBench dataset card · Editorial description reviewed October 29, 2026 |
| Additional source | Model metadata (retrieved October 29, 2026)12 |