The Chinese document tier list nobody benchmarks: vertical classics, tongue diagnosis, waybills, mahjong tiles and industrial QC

Every public multimodal benchmark tests the same things: print, tables, formulas, English handwriting, street-view signage. Not one mainstream benchmark tests vertical classical Chinese, tongue-diagnosis imagery, express waybills, mahjong tiles or industrial QC shots — yet these are what Chinese models process in real businesses every day, and exactly where Western models do not compete at all.

This essay builds a separate mini tier list: eight classes of Chinese documents that nobody tests, assessed across 32 models. It shares only one model with the top three of the 2026 China native multimodal LLM tier list.

Why public benchmarks cannot reach these

Not a technical limitation — three structural reasons:

• The data cannot be public. Waybills carry recipient information, tongue images carry patient information, QC shots carry process parameters. None can enter a public dataset.
• Annotation needs domain experts. The difference between “thin white coating” and “thin yellow coating”, variant-glyph normalisation rules for classical texts, defect grading — annotation costs ten times ordinary OCR or more.
• Overseas institutions have no demand. Vertical traditional texts, red-header document layouts and mahjong tiles have no value to English-world evaluation, so they will never enter MMMU or OCRBench.

Together these mean: on these scenarios there are no public scores to consult; selection depends on vendor vertical cases, scattered community reproductions, and your own acceptance tests. What we publish is a structured desk assessment with the basis marked per item, but it does not replace your 50 images.

The eight classes and their difficulties

Class	Core difficulty	Difference from standard OCR
Vertical classical texts + variant glyphs	Reading order, variant normalisation, unpunctuated segmentation	Right-to-left, top-to-bottom — the reverse of every modern layout
Tongue diagnosis / face reading	Subtle colour transitions, coating texture, lighting bias	Attribute judgement, not character recognition
Express waybills	Handwritten addresses, barcode-text alignment, thermal fading	Handwriting + low-quality medium + strict field structure
Mahjong and tile faces	Fine-grained classes, rotation invariance, occlusion	Dozens of near-identical patterns; one pip is a different tile
Industrial QC defect shots	Hairline scratches, texture suppression, defect grading	The target is “is there, how bad”, not “what does it say”
Chinese menus	Artistic lettering, handwritten price tags, grease glare, dense layout	Recognition + structuring + price inference at once
Red-header documents	Seal occlusion, cross-page seals, document numbers	Seals cover key characters; occlusion recovery is mandatory
Opera masks / artefact patterns	School and era attribution, pattern naming, no text at all	Pure visual knowledge — cultural priors, not OCR

The third column is the point: four of the eight classes (tongue, QC, masks, mahjong) are not OCR tasks at all but visual-knowledge tasks. They appear in an “OCR selection” discussion only because buyers usually procure both capabilities through the same process — a common and expensive mistake.

Per-class results

Each class lists its top three and main failure modes. Scores are 0–100 desk assessments based on vendor vertical cases, comparable published metrics and community reproductions.

Vertical classical texts + variant glyphs

1. GLM-5.3 92 — deepest corpus of official documents and allusions; steadiest variant normalisation
2. Qwen3.5-397B 90 — built on 32-language OCR and document 97; occasional vertical line-order errors
3. DeepSeek-OCR 2 89 — best layout restoration, but outputs the original glyph instead of normalising
4. InternVL 4.0 88 — the best open-source result here; line-order error rate still above the closed-source top three

Failure modes: reading vertical as horizontal (whole-sentence misordering); mixing commentary small text into the body.

Tongue diagnosis / face reading

1. SenseNova 6.5 Pro 86 — most professional-vision and medical vertical cases
2. Huawei Pangu 5.5 84 — industry-model line; hospital delivery experience
3. Spark X2 82 — medical speech and imaging share source corpora

Failure modes: coating-colour judgement shifts wholesale when colour bias is uncorrected; diagnostic rather than descriptive conclusions (a compliance risk).

Express waybills

1. Qwen3.8-Max 96 — document 97; steadiest handwriting-print field alignment
2. DeepSeek-OCR 2 95 — layout specialist; ¥0.8 suits tens of millions per day
3. DeepSeek-V4.1-Flash 91 — excellent on print, weak on handwriting

Failure modes: thermal fading drops trailing digits; province/city inference errors on handwritten addresses.

Mahjong and tile faces

1. Doubao Seed 2.0 Pro 88 — strongest Chinese everyday-life priors
2. MiMo-V2.5 85 — well trained on high-frequency edge scenarios
3. Qwen3.8-Max 83 — strong generalist but thin tile-sample priors

Failure modes: characters/bamboos/circles confuse at low resolution; a half-covered tile gets guessed instead of “undecidable”.

Industrial QC defect shots

1. SenseNova 6.5 Pro 87 — industrial vision is SenseTime’s traditional strength
2. Huawei Pangu 5.5 86 — most production-line deployments; 72-hour fault warning
3. Step-3 81 — good general vision base, lacks vertical fine-tuning

Failure modes: texture background misread as defects (high false positives); cannot grade severity, only answer yes/no.

Chinese menus

1. Doubao Seed 2.0 Pro 95 — dish-name priors far ahead; regional cuisines recognised zero-shot
2. MiMo-V2.5 91 — high-frequency photo-ordering scenarios
3. Spark X2 88 — strong Chinese context; loses points under grease glare

Failure modes: artistic lettering mismatched to handwritten prices (dish A’s price hung on dish B); glare regions left blank.

Red-header documents

1. GLM-5.3 94 — most fluent in official register and layout; stable document-number recognition
2. GLM-5 92 — same-family predecessor; still the public-sector mainstay
3. ERNIE 5.1 89 — Baidu’s government ecosystem

Failure modes: seal-occluded characters skipped rather than recovered; cross-page field misalignment from cross-page seals.

Opera masks / artefact patterns

1. Doubao Seed 2.0 Pro 90 — Chinese culture 97; best heritage and artefact priors
2. SenseNova 6.5 Pro 86 — strong fine-grained vision
3. ERNIE 5.1 84 — cultural knowledge-graph line

Failure modes: mask schools mixed up; patterns named with Western art-history terms instead of Chinese trade vocabulary.

The ninth class: neural-signal decoding, one entrant

The eight classes above all have at least three comparable models. The ninth does not — for EEG/fNIRS temporal plots in neuro-rehabilitation assessment, this site has exactly one entrant: NeuroXess-N1.

That fact matters more than any score. When a class has one entrant:

• There is no comparison to be had; the 88 scenario score for neural-signal decoding only means something against its own previous versions.
• There is no fallback. The procurement decision is not “which model” but “whether to use it”.
• Its evidence grade is C — every figure vendor-reported, with no third-party reproduction. Not because reproduction is hard, but because nobody tried.

Writing classes like this into the mini tier list is the structural difference between this site and every comparison piece: comparisons ask “who is strongest in popular scenarios”; we also record “who is the only one in cold scenarios”. The latter has no traffic but carries information — for the 1% of readers who need it, the page is the only answer on the web.

The mini tier list: only 1 overlap with the overall top three

Weighting the eight classes by business frequency (waybills 20%, menus 15%, red-header 15%, classics 12%, QC 12%, tongue 10%, tiles 8%, masks 8%):

Rank	Model	Weighted	Overall rank	Gap
1	Doubao Seed 2.0 Pro	91.6	2	−1
2	GLM-5.3	90.2	4	−2
3	SenseNova 6.5 Pro	86.4	8	−5
4	Qwen3.8-Max	86.1	1	+3
5	DeepSeek-OCR 2	84.9	20	−14
6	Huawei Pangu 5.5	83.5	21	−14
7	MiMo-V2.5	82.8	7	0
8	Qwen3.5-397B	81.7	3	+5
9	InternVL 4.0	79.8	16	−6

Only Doubao appears in both top threes. Qwen3.8-Max drops from 1st to 4th, Qwen3.5-397B from 3rd to 8th.

Why do the two strongest document models drop out here? Because the Qwen family’s document advantage is built on standard layouts — receipts, tables, cross-page comparison, multilingual — where public benchmarks and large training corpora exist. Tongue, masks, mahjong and QC test cultural priors and vertical case accumulation, and neither is solved by tokenizers or context length.

Conversely, DeepSeek-OCR 2 and Huawei Pangu 5.5 rise 14 places: the former scores high on waybills and classics (layout specialisation), the latter enters the top three on tongue and QC (the industry-model line). They rank 20th and 21st overall and 5th and 6th on cold Chinese documents.

InternVL 4.0’s 79.8 deserves its own reading: it is the only open-source entry on the mini board, 9th, 3.0 points below 8th-place MiMo-V2.5. Open source has caught up to the 90s on standard documents and still trails by a notch on cultural-prior scenarios — that is not a capability gap, it is a gap in how many tongue photographs were in the training data.

This is the rank-divergence index phenomenon recurring in the vertical: the overall rank has near-zero predictive power for specialist models.

The acceptance checklist: how to pick your 50 images

Do not accept-test on public benchmark screenshots. Pick 50 images of your own, in this ratio:

• 20 typical samples (normal lighting, sharp, complete layout)
• 15 boundary samples (faded, glare, occlusion, 15+ degree skew)
• 10 adversarial samples (handwriting mixed with print, seals covering key characters, two similar forms side by side)
• 5 “should refuse” samples (too blurred for a human, key information missing)

The last group is the most overlooked and the most discriminating. Good models say “undecidable”; bad models guess something plausible. On waybills and tongue images, a confident wrong answer costs an order of magnitude more than a refusal.

Record at least three metrics: character accuracy, field-structure restoration rate, and refusal correctness. An evaluation that reports only the first has no reference value.

Why this cycle’s industry models stay out of the eight-class weighting

Two of the six new entries — Guanlan (surveillance video) and Congrong (face recognition) — look like they “should” join this mini board, and the reason they do not is worth stating: the eight classes test “reading Chinese content”, while their home ground is “recognising specific targets” — security defect detection asks “is someone climbing the fence”, face verification asks “is this the same person”, and neither is character recognition or layout understanding. Forcing them into the weighted table would only produce two meaningless low scores.

Their proper homes are two more specialised boards: the surveillance class lives in the security-video field guide (Guanlan is that essay’s deep-dive subject), and recognition plus scenario exclusivity in the industry exclusivity map. The principle of a mini tier list is “one class, one board”: forcing different procurement identities into one weighted table is the composite’s most common mistake.

Extra gates for medical and heritage scenarios

Tongue and face reading are medical settings; capability rank is only a qualifying condition. The real deciding factors are device registration, data-compliance paths and traceability of conclusions — the tier list tells you which models can read a tongue image, not which one may be used to diagnose. Without registration, any model’s output is descriptive reference only.

Heritage and classical texts have a different gate: variant-glyph normalisation must follow specific editorial standards, and a model that emits the original glyph or simplifies on its own creates downstream data-entry trouble. Ask “does it support a specified normalisation table” as a hard requirement instead of reading OCR accuracy.

The full gate checklists for both are in the veto-gate tier list.