Essential cookies keep Instavar working. Optional analytics help us understand how the site is used. Cookie Policy
Manage Cookie Preferences
Service reliability telemetry, including Sentry error monitoring and Vercel Speed Insights, stays enabled so we can secure the product and diagnose failures.
Named-model investigator + Constrained implementerRead bounded results for individual model choices.
DeepSeek earns a narrow lane for blank handling and grounded output, but its measured latency and weak page types argue against a universal default.
Do this first
Test blank handling and grounded output on a small local slice, then decide whether the measured latency fits your resource envelope.
Keep this boundary in view
The 14.8-second figure belongs to the five-model comparison, not the newer four-workflow operational run that measured 17.591 seconds per page.
Inspect the evidence and what would change the conclusion
Evidence layer
Five-model full-50 comparison
Observed
March 2026
Sample
50 pages across seven types; 49 processed by DeepSeek
What supports it
In the five-model 50-page comparison it processed 49 pages, detected all three blank pages, and averaged 14.8 seconds per page.
Change our mind if
On a representative local set, DeepSeek loses its blank or grounding advantage, or meets all quality and speed needs well enough to become the default.
Claim OCR-DEEPSEEK-50-V1 / v1, checked 6 Aug 2026
This post answers a narrow production question: where does DeepSeek OCR-2 belong if its aggregate benchmark score is not the best.
The short version
DeepSeek OCR-2 is a solid markdown-oriented OCR workflow with the best blank-page detection in our benchmark.
Its aggregate CER was 39.34%, placing it 4th of 5 models in the full-50 comparison.
It was also the slowest path in this five-model comparison at 14.8 s/page.
Use it when blank detection or grounded output matters.
Avoid it as the default OCR model for formulas, worksheets, degraded scans, or high-volume processing.
The one-minute decision path
The aggregate score makes DeepSeek OCR-2 look mediocre. The page-type breakdown explains why it still matters.
Its value is not broad accuracy. Its value is operational hygiene: it avoids blank-page hallucination and gives grounded output when you need to trace extracted text back to the page.
If your bottleneck is...
DeepSeek fit
Better first test
blank-page detection
strong fit
DeepSeek, Hunyuan, or Qianfan
grounded output with coordinates
useful fallback
Hunyuan first, then DeepSeek
text-first notes
acceptable
Qianfan or Hunyuan
formulas, worksheets, or degraded scans
Local test / OCR-DEEPSEEK-50-V1
Test this claim on your documents.
Copy a read-only experiment brief that asks your coding agent to preserve raw evidence, test a negative case, and return a verdict that can disagree with this page.
Use Codex to test this bounded OCR claim in my local environment.
Source: https://instavar.com/research/ocr/deepseek-ocr-production-notes-2026
Claim ID: OCR-DEEPSEEK-50-V1
Claim version: v1, checked 6 Aug 2026
Claim: DeepSeek earns a narrow lane for blank handling and grounded output, but its measured latency and weak page types argue against a universal default.
Observed evidence: In the five-model 50-page comparison it processed 49 pages, detected all three blank pages, and averaged 14.8 seconds per page.
Evidence layer: Five-model full-50 comparison.
Sample: 50 pages across seven types; 49 processed by DeepSeek.
Boundary: The 14.8-second figure belongs to the five-model comparison, not the newer four-workflow operational run that measured 17.591 seconds per page.
Potential falsifier: On a representative local set, DeepSeek loses its blank or grounding advantage, or meets all quality and speed needs well enough to become the default.
Safety and privacy:
- Work read-only by default and keep all documents and outputs local.
- Do not upload documents, reveal credentials, or send document contents to an external service.
- Do not install software, download large models, start paid jobs, or change my environment without asking first.
- If the available tools cannot run the comparison safely, stop with a concrete setup plan instead of pretending the claim was tested.
Test method:
1. Inspect the available OCR workflows and record their exact versions, runtime, hardware, and configuration.
2. Help me choose a small representative sample. Include a known difficult page and a negative control with expected text or structure when available.
3. Before running anything, state the primary failure we are testing and the measurement that would expose it.
4. Preserve raw outputs. Compare text fidelity, structure, blank-page behavior, grounding correctness, runtime, and cleanup effort only where those properties can actually be measured.
5. Treat anchor count as an unverified quantity until sampled anchors are checked against the page.
6. Report the result as support, contradict, mixed, or inconclusive. Explain what observation produced that verdict.
7. Name the smallest follow-up test that could change the verdict.
weak fit
Qianfan or GLM, depending on page type
high-volume batch processing
poor fit
faster service-backed lanes
Where this fits
For founders: DeepSeek OCR-2 has a narrow but real role. Use it when blank pages and page coordinates matter. Do not use it as the default OCR model for every scanned PDF.
For engineers: Use this page to understand where DeepSeek fits in a multi-model routing pipeline, what its failure modes look like on degraded scans, and when to route pages away from it.
DeepSeek OCR-2 is a vision-language model (VLM) from DeepSeek designed for document OCR. Like several other models in the current open-source OCR wave, it combines a vision encoder with a language decoder to produce structured text output from page images.
If that sentence is too abstract: it reads the page image, then writes the extracted document text. Unlike older OCR tools, it can also return structure such as markdown blocks and page coordinates.
Two properties set it apart from the other models in our benchmark:
Markdown-oriented output. DeepSeek produces clean, structured markdown by default. Headings, lists, and tables come through with reasonable fidelity, which reduces downstream cleanup cost when the output feeds into a markdown-native pipeline.
Grounded output with bounding boxes. DeepSeek supports coordinate grounding - it can return bounding box positions alongside extracted text. In our benchmark, it was the second-best grounded workflow after Hunyuan.
2 Benchmark scores vs reality
The aggregate numbers for DeepSeek look mediocre at first glance:
Metric
DeepSeek OCR-2
Average CER
39.34%
Average WER
33.39%
Pages processed
49
Speed
14.8 s/page
A 39.34% CER puts DeepSeek 4th out of 5 models in our full-50 benchmark. That is not a competitive aggregate number.
But the aggregate hides the interesting finding. DeepSeek's errors are not evenly distributed across page types. It is a poor fit for some pages and competitive on others.
Plain-English read: DeepSeek is not the best all-purpose choice. Its value is that it handles blank pages well and gives you coordinates when you need to trace extracted text back to the page.
The same caution shows up in community evaluation threads. In the DeepSeek OCR-2 release discussion, users compared DeepSeek against PaddleOCR, Mistral, and OCR Arena rather than accepting one reported score. In the earlier DeepSeek OCR release thread, users asked for OmniDocBench details and noted that edit distance alone does not explain table or formula behavior. In a PaddleOCR-VL-1.5 thread, users reported that strong benchmark scores can still coexist with simple table failures, repetition errors, CPU slowness, and hardware questions.
That is why this page treats DeepSeek as a routing lane, not a leaderboard winner. If a pipeline needs blank-page hygiene or grounded coordinates, DeepSeek deserves a test. If the pipeline needs formulas, worksheets, degraded scans, or high-volume throughput, the benchmark table below says to route away.
CER breakdown by page type
Page type
Pages
DeepSeek CER
Best model
Best CER
text_first_notes
10
8.3%
Qianfan
5.9%
diagram_question
10
30.2%
GLM
6.1%
formula_heavy
8
76.6%
Qianfan
20.7%
table_heavy
8
43.8%
Qianfan
15.7%
worksheet_options
8
46.5%
Qianfan
7.1%
low_contrast_or_faint_scan
3
69.1%
Qianfan
0.0%
blank_or_near_blank
2
0.0%
DeepSeek/Hunyuan/Qianfan
0.0%
The 8.3% CER on text-first notes is within striking distance of the best model. The 0.0% on blank pages is perfect. Everything else is a weakness.
DeepSeek detected all 3 blank pages in our 50-page corpus with 0.0% CER. This sounds trivial, but blank detection is a real pipeline hygiene problem. Models that hallucinate text on blank pages inject noise into downstream processing, create phantom entries in document indices, and waste compute on non-content.
GLM, for comparison, failed all 3 blank pages - it hallucinated content on every one.
Only three models in our benchmark achieved 3/3 blank detection: DeepSeek, Hunyuan, and Qianfan.
3.2 Text-first notes
On the text_first_notes page type (10 pages of handwritten and printed notes, bullets, and worked answers), DeepSeek scored 8.3% CER. Qianfan led at 5.9%, but DeepSeek's result is competitive and usable in production without heavy post-processing.
3.3 Grounded output
DeepSeek is the second-best grounded workflow in our benchmark, after Hunyuan. If your pipeline needs bounding box coordinates alongside extracted text - for anchor overlays, region-level confidence scoring, or spatial search - DeepSeek is one of only two models that deliver this reliably.
3.4 Markdown output quality
The markdown DeepSeek produces is structurally clean. Tables render correctly, heading hierarchy is preserved, and list formatting is consistent. For pipelines that consume markdown directly (rendering, indexing, downstream LLM input), this reduces the cleanup step.
4 Where it fails
4.1 Low-contrast scans
CER of 69.1% on low-contrast and faint scans. Qianfan gets 0.0% on the same pages. Hunyuan gets 6.6%. DeepSeek's performance on degraded input is not competitive.
If your corpus includes photocopied worksheets, faded thermal prints, or low-DPI scans, route these pages away from DeepSeek.
4.2 Worksheets
46.5% CER on worksheet_options pages (multiple-choice layouts, grid-style answers). Qianfan gets 7.1% on the same pages. The gap is large enough that DeepSeek should not be used as the primary model for worksheet-heavy documents.
4.3 Formulas
76.6% CER on formula_heavy pages. This is the worst result among DeepSeek's page types. Qianfan leads at 20.7%, and even that is not great - formula OCR remains hard across all models. But DeepSeek's result here is unusable without significant post-correction.
4.4 Diagrams
30.2% CER on diagram_question pages, where GLM leads at 6.1%. DeepSeek does not preserve question-local visuals or diagram-linked text as reliably as GLM.
5 Speed and cost
At 14.8 seconds per page, DeepSeek is the slowest model in this five-model comparison by a wide margin. A newer four-workflow operational run measured 17.591 s/page; that result uses a different workflow boundary and is reported separately on the leaderboard.
Model
Speed
GLM
0.9 s/page
DeepSeek
14.8 s/page
GLM is roughly 16x faster. A 1,000-page document takes DeepSeek over 4 hours. The same document takes GLM about 15 minutes.
This makes DeepSeek impractical for high-volume processing. It is only viable for small-batch workflows or quality-critical lanes where its specific strengths (blank detection, grounding) justify the latency cost.
6 Comparison: DeepSeek vs GLM vs Qianfan
These are the three models most commonly compared for markdown-oriented OCR in our pipeline. Here is the head-to-head:
Metric
DeepSeek
GLM
Qianfan
Avg CER
39.34%
33.84%
12.80%
Avg WER
33.39%
27.59%
13.18%
Speed
14.8 s/page
0.9 s/page
N/A
Blank detection
3/3 (100%)
0/3 (failed)
3/3 (100%)
Best page type
blank_or_near_blank, text notes
diagram_question
5 of 7 page types
Qianfan wins 5 of 7 page types outright. GLM wins diagrams and is dramatically faster. DeepSeek's wins are narrower: blank detection (shared with Qianfan) and grounded output (second to Hunyuan).
The honest read: if you do not need grounded output or bounding box coordinates, Qianfan is the stronger default and GLM is the faster one. DeepSeek earns its lane only when its specific strengths matter to your pipeline.
6.1 DeepSeek vs GLM
Use DeepSeek OCR-2 when blank-page detection or grounded coordinates are the requirement. It detected all blank pages in our corpus and gives a stronger coordinate path than GLM.
Use GLM-OCR when speed or diagram-linked question pages matter more. GLM was 16x faster in our run and led the diagram_question page type at 6.1% CER. The tradeoff is blank-page hygiene: GLM failed all blank pages in this corpus, so do not use it alone on batch-scanned documents with separators or empty backs.
6.2 DeepSeek vs PaddleOCR
Use DeepSeek OCR-2 when you specifically need blank handling, grounding, or a DeepSeek-compatible markdown lane.
Keep PaddleOCR-VL-1.5 in the shortlist when ecosystem depth and local deployment patterns matter. Reddit users testing PaddleOCR-VL care about table behavior, hardware support, and whether benchmark scores survive real pages. That is the right framing: compare it against DeepSeek on your table-heavy, formula-heavy, and degraded pages rather than using either model as a universal default.
7 When to use DeepSeek OCR-2
Route to DeepSeek when:
Blank detection matters. If your pipeline processes mixed documents with intermittent blank or near-blank pages, DeepSeek's 3/3 detection prevents phantom entries.
You need grounded coordinates. For anchor overlays, spatial search, or region-level extraction, DeepSeek is the second-best grounded workflow.
The pages are text-first. On notes, bullets, and worked answers, DeepSeek's 8.3% CER is competitive.
Route away from DeepSeek when:
Formulas are present. 76.6% CER is not salvageable.
Worksheets dominate. 46.5% CER - use Qianfan instead.
Scans are degraded. Low-contrast, faded, or faint pages break DeepSeek badly.
Volume is high. The five-model comparison measured 14.8 s/page, which makes the tested path impractical for batch processing above a few hundred pages.
Default alternative
Qianfan wins 5 of 7 page types in our benchmark and has competitive blank detection. If you need a single markdown-oriented OCR model and do not require grounded output, start with Qianfan.
No. In our 50-page benchmark, it places 4th of 5 models by aggregate CER. It has specific strengths (blank detection, grounded output, text notes) but is not the strongest general-purpose choice.
Can I use DeepSeek OCR-2 as my only OCR model?
You can, but you will get poor results on formulas, worksheets, and low-contrast scans. A multi-model routing pipeline that sends different page types to different models will outperform any single-model deployment.
How does DeepSeek compare to Hunyuan for grounded output?
Hunyuan is the strongest grounded workflow in our benchmark. DeepSeek is second. If grounded output is your primary requirement, test Hunyuan first. If you also need blank detection, DeepSeek adds value as a complementary lane.
Why is DeepSeek so slow?
At 14.8 s/page, DeepSeek was the slowest path in the five-model comparison. GLM measured 0.9 s/page in that same run, or roughly 16x faster. Those observations establish a workflow-level latency difference in this setup. They do not isolate model architecture as the cause, and they should not be combined with the newer four-workflow timings.
Should I use DeepSeek or Qianfan for markdown OCR?
Qianfan wins 5 of 7 page types and has lower aggregate CER (12.80% vs 39.34%). Unless you specifically need grounded output or bounding box coordinates, Qianfan is the better markdown-oriented default.