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.
Hunyuan's grounding helps on some faint, formula, and blank pages, while FireRed's Markdown-first path is faster and stronger on several structured page types.
Do this first
Separate faint, formula, blank, and structured pages, then compare quality and runtime by page type instead of averaging them together.
Keep this boundary in view
Cross-model consensus is a practical reference, not manually transcribed ground truth, and neither model wins every page type.
Inspect the evidence and what would change the conclusion
Evidence layer
Hunyuan and FireRed full-50 comparison
Observed
March 2026
Sample
50 pages across seven types; 49 processed per workflow
What supports it
The comparison used the same 50-page, seven-type slice; each workflow processed 49 pages.
Change our mind if
A manually adjudicated held-out set removes the observed page-type split or reverses the operational trade-off.
Claim OCR-HUNYUAN-FIRERED-50-V1 / v1, checked 6 Aug 2026
This post answers a head-to-head question: if you are choosing between Hunyuan OCR and FireRed OCR, which one should handle which scanned pages.
The short version
Neither Hunyuan OCR nor FireRed OCR is universally better.
Across 50 scanned pages in 7 page types, Hunyuan won 4 page types and FireRed won 3.
Hunyuan is stronger when the workflow needs coordinate-grounded output or messy-page reasoning.
FireRed is faster and cleaner when the page already has clear structure.
The practical answer is to route by page type, not pick one model globally.
The one-minute decision path
The split is architectural. Hunyuan returns text with page coordinates, so it can help on degraded, ambiguous, or blank pages where spatial evidence matters. FireRed produces markdown-first output, so it is faster and cleaner on structured content that already reads clearly.
If the page is...
Start with...
Why
low-contrast, faint, blank, or formula-heavy
Hunyuan OCR
coordinate-grounded output gives better evidence on harder pages
table-heavy, worksheet-style, or diagram-question
FireRed OCR
markdown-first output was cleaner or faster in these slices
text-first notes
either, then optimize for speed or cleanup cost
the measured gap was small
Local test / OCR-HUNYUAN-FIRERED-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/hunyuan-ocr-vs-firered-ocr-2026
Claim ID: OCR-HUNYUAN-FIRERED-50-V1
Claim version: v1, checked 6 Aug 2026
Claim: Hunyuan's grounding helps on some faint, formula, and blank pages, while FireRed's Markdown-first path is faster and stronger on several structured page types.
Observed evidence: The comparison used the same 50-page, seven-type slice; each workflow processed 49 pages.
Evidence layer: Hunyuan and FireRed full-50 comparison.
Sample: 50 pages across seven types; 49 processed per workflow.
Boundary: Cross-model consensus is a practical reference, not manually transcribed ground truth, and neither model wins every page type.
Potential falsifier: A manually adjudicated held-out set removes the observed page-type split or reverses the operational trade-off.
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.
mixed scanned document
route by page type
either single-model choice loses on several slices
Where this fits
For founders: If you are choosing between Hunyuan and FireRed, the answer is "both, routed by page type." A single-model stack will underperform on at least 3 of the 7 page types we tested.
For engineers: Use the CER breakdown by page type to set routing thresholds. FireRed at 3.4 s/page is the speed default; Hunyuan at 6.6 s/page is the fallback for degraded inputs. Never send blank-detection jobs to FireRed - it hallucinates phantom text.
Architecture comparison
Hunyuan and FireRed take fundamentally different approaches to document understanding.
Hunyuan OCR produces coordinate-grounded output. In plain terms, it returns text together with page coordinates. Every text region comes with bounding boxes, giving downstream systems spatial context for where text lives on the page. This grounding is especially useful for degraded scans where the model needs to distinguish real content from noise, and for blank-page detection where it can confirm the absence of text regions rather than guessing.
FireRed OCR produces markdown-first output. The model directly generates structured Markdown - headings, tables, lists - without an intermediate spatial representation. This makes it faster (no bounding-box overhead) and produces cleaner output on pages that already have clear visual structure. The tradeoff is that it lacks spatial grounding, which hurts on ambiguous or degraded inputs.
Dimension
Hunyuan OCR
FireRed OCR
Output format
Coordinate-grounded (bounding boxes)
Markdown-first
Processing speed
6.6 s/page
3.4 s/page
Strength
Degraded scans, spatial reasoning
Structured content, speed
Weakness
Slower, struggles with complex tables
Blank-page hallucination, formula parsing
Test setup
We evaluated both models on 50 scanned pages drawn from 7 page types: text-first notes, diagram questions, formula-heavy pages, table-heavy pages, worksheet/options pages, low-contrast or faint scans, and blank or near-blank pages.
Evaluation used Character Error Rate (CER) and Word Error Rate (WER) computed via cross-model consensus - the same framework described in our benchmark methodology. Cross-model consensus means we compared several OCR outputs and used majority agreement as the practical reference. Ground truth was established through majority agreement across 5 OCR models, not manual transcription.
Both models processed 49 of the 50 pages (one page was excluded from each run due to processing failures). All runs used the same input images with no preprocessing differences.
How to read this comparison: the overall score tells you the average. The page-type table tells you what to do in production. For OCR, the second table is usually more useful.
The aggregate numbers are deceptively close. The real story is in the page-type breakdown.
Head-to-head results by page type
CER summary by page type
Page type
Pages
FireRed CER%
Hunyuan CER%
Winner
text_first_notes
10
10.0
8.2
Hunyuan
diagram_question
10
39.9
65.9
FireRed
formula_heavy
8
78.7
42.5
Hunyuan
table_heavy
8
39.7
63.6
FireRed
worksheet_options
8
12.2
16.2
FireRed
low_contrast_or_faint_scan
3
16.3
6.6
Hunyuan
blank_or_near_blank
2
158.8
0.0
Hunyuan
Text-first notes
Winner: Hunyuan (8.2% vs 10.0%)
Both models handle clean, text-heavy pages competently. Hunyuan edges ahead by a small margin. At these error rates, the practical difference is minimal for most downstream tasks. If speed matters more than a 1.8 percentage-point CER gap, FireRed is the reasonable choice here.
Diagram questions
Winner: FireRed (39.9% vs 65.9%)
FireRed wins by a significant margin on pages with inline diagrams. Hunyuan's coordinate grounding does not translate into better diagram comprehension - the bounding-box approach appears to fragment diagram-adjacent text, inflating errors. FireRed's markdown pipeline handles the mix of text and visual elements more cleanly.
Formulas
Winner: Hunyuan (42.5% vs 78.7%)
This is Hunyuan's strongest page-type win. FireRed's markdown pipeline struggles with mathematical notation - LaTeX-style expressions are frequently malformed or truncated in its output. Hunyuan's spatial grounding helps it preserve formula structure, though a 42.5% CER still means significant errors. Neither model is production-ready for formula-heavy pages without post-processing.
Tables
Winner: FireRed (39.7% vs 63.6%)
Markdown table extraction is FireRed's architectural strength. Its output format naturally maps to table structures, while Hunyuan's coordinate-based approach must reconstruct table relationships from spatial positions - a harder problem. For table-heavy documents, FireRed is the clear default.
Worksheets
Winner: FireRed (12.2% vs 16.2%)
Both models handle structured worksheets with multiple-choice options well, producing relatively low error rates. FireRed has a modest edge. The structured, repetitive layout of worksheets plays to the strengths of markdown-first output.
Low-contrast scans
Winner: Hunyuan (6.6% vs 16.3%)
Hunyuan wins decisively on faded or low-contrast scans. Coordinate grounding helps here: the model can locate text regions even when contrast is poor, while FireRed's pipeline is more likely to miss or misread faint characters. For document processing pipelines that handle aged or photocopied material, this result matters.
Blank pages
Winner: Hunyuan (0.0% vs 158.8%)
This is the most critical finding in the benchmark. Hunyuan correctly identifies blank pages and returns empty output. FireRed generates phantom text on blank pages - hallucinated content that does not exist in the source image. A CER of 158.8% means FireRed produces more characters of hallucinated text than a typical page contains.
This is not a minor edge case. Any document processing pipeline encounters blank pages - separator sheets, blank backs of single-sided prints, intentional empty pages. A model that hallucinates content on blank pages will inject noise into every downstream system.
Speed comparison
FireRed processes pages at 3.4 seconds per page - nearly twice as fast as Hunyuan's 6.6 seconds per page. For a 100-page document, that is the difference between roughly 6 minutes and 11 minutes of processing time.
The speed gap is architectural. FireRed's markdown-first pipeline skips bounding-box computation, while Hunyuan must compute spatial coordinates for every detected text region before generating output.
For latency-sensitive applications (real-time document intake, user-facing OCR), FireRed's speed advantage is significant. For batch processing where accuracy matters more than throughput, Hunyuan's slower pace is an acceptable tradeoff.
When to use which
Document type
Use Hunyuan when...
Use FireRed when...
Text notes
Slight accuracy edge needed
Speed matters more
Diagrams
Avoid
Default choice
Formulas
Default choice
Avoid
Tables
Avoid
Default choice
Worksheets
Acceptable
Default choice
Low-contrast scans
Default choice
Acceptable fallback
Blank detection
Default choice
NEVER (hallucination risk)
The routing logic in practice: classify incoming pages by page type (using a lightweight vision classifier or heuristic rules), then route to the appropriate model. FireRed is the speed-optimised default for structured content. Hunyuan is the accuracy fallback for degraded inputs, formulas, and blank-page gating.
A blank-detection step before the main OCR pass is worth implementing regardless of which model you use for content extraction. Hunyuan's perfect blank-page detection makes it a strong candidate for this gating role.
FAQ
Is one model strictly better than the other?
No. Hunyuan wins 4 of 7 page types and FireRed wins 3 of 7. The wins depend on page type and are driven by architectural differences, not overall model quality. A routed approach using both models outperforms either one alone.
Can I use FireRed for everything if I need speed?
You can, with one hard exception: never use FireRed for blank-page detection. Its hallucination on blank pages (158.8% CER) will inject phantom content into your pipeline. For all other page types, FireRed is a reasonable single-model choice if you accept higher error rates on formulas and low-contrast scans.
How does cross-model consensus work as ground truth?
We use majority agreement across 5 OCR models as the reference output instead of manual human transcription. This approach scales better than manual annotation and removes single-model bias. The full methodology is documented in our benchmark methodology guide.
What about other models like GLM-OCR or DeepSeek?
This post focuses on the Hunyuan vs FireRed comparison. For the broader model landscape including GLM-OCR, DeepSeek-OCR-2, dots.ocr-1.5, and PaddleOCR-VL-1.5, see the OCR Model Leaderboard and the workflow-fit guide.
Should I run both models on every page?
Running both models on every page doubles your compute cost and latency. A better approach is to classify pages first and route to the stronger model for that page type. The exception is high-value documents where you want consensus - running both and comparing output can catch errors that either model alone would miss.