DeepSeekLanguage model
DeepSeek OCR 2
DeepSeek OCR 2 has no overall index value yet because there are too few independent benchmark results.
PopularityRank 263 of 785
41out of 100-1 vs. last month
based on search, usage and attention
Summary
What is distinctive about DeepSeek OCR 2 is its document processing with dynamic resolution. The 3-billion-parameter checkpoint combines multiple crops with an overview and, according to its model card, supports free OCR and document-to-Markdown prompts.
DeepSeek OCR 2 is limited to layouts and text recognition rather than general chat or tool loops. Its model card documents local serving, Apache 2.0, and an 8,192-token context for these OCR pipelines.
All benchmark results
Every published result with its source, test setup and comparison models.
Show all 89 published benchmark results
Page 1 of 8
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.
OmniDocBench v1.5 (Show measurement, test conditions, and source)
- Source value
- 91.09
- Score
- 91.09
- Metric
- Overall document-reading score
- Unit
- %
- Benchmark version
- DeepSeek-OCR 2 arXiv v1, Table 1 (OmniDocBench v1.5)
- Category
- Overall
- Direction
- Higher is better
- Sample
- 1355
- Source type
- vendor-reported
- Evaluator
- DeepSeek-AI, DeepSeek-OCR 2: Visual Causal Flow (arXiv v1)
- Status
- active
- Retrieved at
- 2026-10-17
- Methodology
- Vendor-reported result. Exact DeepSeek-OCR 2 row in Table 1, aggregate document-reading result on OmniDocBench v1.5. The benchmark has 1,355 document pages across nine categories in Chinese and English; the paper caps visual tokens at 1,120 per page.
- Context
- The paper does not publish a model-specific inference route, immutable serving revision, or full prompt and decoding settings.
DeepSeek OCR 2 production repetition rate, online-user-logs (image) (Show measurement, test conditions, and source)
- Source value
- 4.17
- Score
- 4.17
- Metric
- Repetition rate, lower is better
- Unit
- %
- Benchmark version
- DeepSeek-OCR 2 technical report arXiv:2601.20552v1, Table 4 production comparison
- Category
- production OCR quality proxy
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek-AI, DeepSeek-OCR 2: Visual Causal Flow
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Vendor-reported result. The paper evaluates the exact DeepSeek-OCR 2 model in an online OCR service reading user-log images and documents for DeepSeek LLMs. This is the source's production condition "online-user-logs (image)". It states that ground truth is unavailable in production and reports repetition rate as the primary observable quality metric. It does not disclose the model serving revision, sample denominator, rate formula or confidence interval.
- Context
- This is a vendor-reported production-system metric, not held-out OCR accuracy. The paper says production ground truth is unavailable and does not publish a cell denominator or confidence interval; repetition rate is its primary observable quality metric.
DeepSeek OCR 2 production repetition rate, pretrain-data (PDF) (Show measurement, test conditions, and source)
- Source value
- 2.88
- Score
- 2.88
- Metric
- Repetition rate, lower is better
- Unit
- %
- Benchmark version
- DeepSeek-OCR 2 technical report arXiv:2601.20552v1, Table 4 production comparison
- Category
- production OCR quality proxy
- Direction
- Lower is better
- Source type
- vendor-reported
- Evaluator
- DeepSeek-AI, DeepSeek-OCR 2: Visual Causal Flow
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Vendor-reported result. The paper evaluates the exact DeepSeek-OCR 2 model in a batch PDF-processing pipeline for pretraining data. This is the source's production condition "pretrain-data (PDF)". It states that ground truth is unavailable in production and reports repetition rate as the primary observable quality metric. It does not disclose the model serving revision, sample denominator, rate formula or confidence interval.
- Context
- This is a vendor-reported production-system metric, not held-out OCR accuracy. The paper says production ground truth is unavailable and does not publish a cell denominator or confidence interval; repetition rate is its primary observable quality metric.
GlotOCR Bench high-resource tier (Show measurement, test conditions, and source)
- Source value
- 50.1
- Score
- 50.1
- Metric
- A@0
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench high-resource tier (Show measurement, test conditions, and source)
- Source value
- 76.2
- Score
- 76.2
- Metric
- A@5
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench mid-resource tier (Show measurement, test conditions, and source)
- Source value
- 22.2
- Score
- 22.2
- Metric
- A@0
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench mid-resource tier (Show measurement, test conditions, and source)
- Source value
- 39.7
- Score
- 39.7
- Metric
- A@5
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench low-resource tier (Show measurement, test conditions, and source)
- Source value
- 0.1
- Score
- 0.1
- Metric
- A@0
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench low-resource tier (Show measurement, test conditions, and source)
- Source value
- 0.3
- Score
- 0.3
- Metric
- A@5
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench across-tier mean (Show measurement, test conditions, and source)
- Source value
- 24.1
- Score
- 24.1
- Metric
- A@0
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
GlotOCR Bench across-tier mean (Show measurement, test conditions, and source)
- Source value
- 38.7
- Score
- 38.7
- Metric
- A@5
- Unit
- %
- Benchmark version
- GlotOCR Bench, arXiv:2604.12978v1, Table 2
- Category
- OCR
- Direction
- Higher is better
- Source type
- community-reported
- Evaluator
- GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Community-reported result. Zero-shot evaluation with the authors' uv-scripts/ocr inference suite. A chat template is applied when available; otherwise the prompt is passed directly. The prompt requests transcription wrapped in tags with no commentary. Images use native rendered resolution without preprocessing. Table 2 reports per-tier macro averages across scripts: high-resource has 1 script, mid-resource 9, and low-resource 148. The Mean columns are the reported average across tiers, not a pooled instance score. CER is the best value among original, reversed, lowercased, and Unicode-combining-mark-removed output configurations. The source does not state per-cell example counts or confidence intervals.
- Context
- This is a primary academic benchmark paper's result, not a provider-reported score. The authors' affiliations do not by themselves establish a formal independent third-party evaluation. The table identifies DeepSeek-OCR-2 by name but does not give an evaluated checkpoint hash or serving revision. Script counts are not per-cell example counts; the paper does not report per-cell N or confidence intervals.
PureDocBench Legal_Gov, Clean, FormulaCDM (FCM) (Show measurement, test conditions, and source)
- Source value
- 45.9
- Score
- 45.9
- Metric
- FormulaCDM (FCM); Table 9 marks higher as better (↑), and the abstract reports the formula metric as a percentage.
- Unit
- %
- Benchmark version
- PureDocBench, May 2026 preprint arXiv:2605.07492, Table 9
- Direction
- Higher is better
- Source type
- independent-evaluator
- Evaluator
- PureDocBench paper, arXiv:2605.07492 preprint (Li et al.)
- Status
- active
- Retrieved at
- 2026-10-18
- Methodology
- Independent evaluation. Exact source row label “DeepSeek-OCR-2 3B” in PureDocBench Table 9 for Legal_Gov (151 pages). The paper describes PureDocBench as 1,475 pages across 10 domains and 66 subcategories, each represented in clean, digital-degraded, and real-degraded versions. This cell is for the stated track. The table does not identify a pinned model revision, quantization, decoding parameters, or per-cell denominator. Track: Clean.
- Context
- This is a May 2026 preprint; the hosting page warns that preprints may not have been peer reviewed. No sample count or confidence interval is stated for this exact row. Units remain unknown where not defined by the paper.
Page 1 of 8
Popularity compared
How often the model is searched for, used and mentioned, compared with every model of its category.
Popularity index
The 100 most popular models of this category, 0 to 100, higher is more popular
20 of 101 models
As of November 2026 · Source: gradually.ai popularity index
Over time
Weekly values from October 18, 2026 to November 7, 2026
DeepSeek OCR 2Ornith 1.0 9B 41Phi-4-multimodal-instruct 41Granite 4.2 30B 41DeepSeek V4 Pro 0813 41
Method changes in the timeline, a jump at these points is no market event:
- 1November 1, 2026: Value only with search or usage data, 30-day attention
What is gaining popularity right now, across models, vendors, harnesses and AI subscriptions, is collected in All AI trends.
Similar models
More models of the same kind as DeepSeek OCR 2, chosen by model line, developer, size, price and release date.
- DeepSeek R1 0528 Qwen3 8BDeepSeek · same developer
- DeepSeek-V4-FlashDeepSeek · same developer
- DeepSeek-V4-ProDeepSeek · same developer
- DeepSeek-V3.2DeepSeek · same developer
- DeepSeek-V3.2 ExpDeepSeek · same developer
- DeepSeek V4.1 Flash 1017DeepSeek · same developer
- DeepSeek-V3.1 TerminusDeepSeek · same developer
- DeepSeek-V4.1-FlashDeepSeek · same developer
All specifications
Architecture
Input and output
Watermarking and provenance
AI watermarking refers to invisible signals embedded in the output. Visible marks and provenance metadata are listed separately. Unknown means there is insufficient evidence for that particular field.
- AI watermarking
- Unknown
- no source
- Watermarking technique
- Not known
- no source
Licence and access
Frequently asked questions about DeepSeek OCR 2
How large is the context window of DeepSeek OCR 2?
8,192 tokens. That is how much text DeepSeek OCR 2 reads at once, about 10 A4 pages.
Page count as a rule of thumb from Artificial Analysis
GLM-5.1200K
DeepSeek-V3 0324164K
GLM-4.5131K
GPT OSS 120B131K
DeepSeek-V3.2128K
DeepSeek-R1128K
DeepSeek OCR 2Rank 103 of 1038K
Can I run DeepSeek OCR 2 locally?
Yes, the weights are open.
Stated by DeepSeek
- Open weights
- yes
- License
- Apache 2.0
- Parameters
- 3B
- active per token
- n/a
Sources and data date
Show all 22 sources
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 |
|---|---|
| Input | Documented modalities. Retrieved October 27, 2026. |
| Output | Documented modalities. Retrieved October 27, 2026. |
| Research status | Research date October 15, 2026. 14 source URLs checked. This documents the inspected sources, not an exhaustive inventory of every publication. |
| Not published | ocr-leaderboard-protocol Note Independent widget results are published, but exact OCR preprocessing, image resolution, decoding settings, and task/run dates are not exposed in the checked source. |
| Not published | immutable served model revision and exact endpoint Note The added source row names the model but does not publish an immutable served revision or, where applicable, exact provider endpoint. No result is transferred to a dated sibling variant. |
| Not published | unreported uncertainty metadata Note Unreported confidence intervals and per-subcategory sample counts remain absent; displayed ± terms with undefined meaning are not converted to intervals. |
| Additional source | Hugging Face (retrieved October 15, 2026) · DeepSeek OCR 2 model card · Editorial description reviewed October 2, 2026 |
| Additional source | Allen Institute olmOCR-bench leaderboard (retrieved October 15, 2026) |
| Additional source | PaddlePaddle Real5-OmniDocBench leaderboard (retrieved October 15, 2026) |
| Additional source | DeepSeek-AI, DeepSeek-OCR 2: Visual Causal Flow (arXiv v1) (retrieved October 17, 2026; October 18, 2026) |
| Additional source | ParseBench official leaderboard (retrieved October 17, 2026) |
| Additional source | GlotOCR Bench: OCR Models Still Struggle Beyond a Handful of Unicode Scripts (retrieved October 18, 2026) |
| Additional source | Infinity-Parser2 Technical Report (arXiv:2607.07836) (retrieved October 18, 2026) |
| Additional source | FireRed-OCR Technical Report v1 (arXiv:2603.01840) (retrieved October 18, 2026) |
| Additional source | PureDocBench paper, arXiv:2605.07492 preprint (Li et al.) (retrieved October 18, 2026) |
| Additional source | PureDocBench official leaderboard (retrieved October 18, 2026) |
| Additional source | Model metadata (retrieved October 17, 2026)1234 |
| Additional source | DeepSeek OCR 2 paper · Editorial description reviewed October 2, 2026 |
| Additional source | OpenRouter: Popularity index (derived), CC BY 4.0 |
| Additional source | Vercel AI Gateway: Popularity index (derived), CC BY 4.0 |
| Additional source | Ollama: cumulative downloads for the popularity index |
| Additional source | Replicate: cumulative runs for the popularity index |
| Additional source | Google Trending Now: limited feed selection for Germany and the US |
| Additional source | Creative Commons Attribution 4.0 |