Skip to content
CN Multimodal Tier List2026 · 8 DIMENSIONS
中文

Method · weights · evidence

How this tier list is calculated

We do not claim a closed-lab replication. This is structured scoring from public evidence: third-party boards, official reports, community reproductions and scenario checklists mapped onto eight dimensions, with uncertainty made explicit.

8dimensions
14grade A
4score bands

In one sentence

Composite = Σ (dimension score × weight) out of 100, with weights 20/18/14/14/14/6/6/8 for perception, documents, charts, video, agents, generation, Chinese and ecosystem. The composite is for ranking only — use dimension scores for selection. Every model carries an A/B/C evidence grade; only A means third-party cross-verification.

Formula

A simple weighted sum; the hard part is evidence

composite = 0.20·P + 0.18·D + 0.14·C + 0.14·V + 0.14·A + 0.06·G + 0.06·Z + 0.08·E

P perception · D documents · C charts · V video · A agents · G generation · Z Chinese · E ecosystem. Each 0–100, one decimal.

DimensionWeightDefinition and what a high score means
Visual perception & grounding20%Complex-scene recognition, fine-grained attributes, spatial relations, referring grounding (bbox/point) and Chinese UI elements.
Document parsing & Chinese OCR18%Receipts, tables, handwriting, vertical text, formulas and low-quality scans; character accuracy plus layout fidelity.
Chart & document reasoning14%Data extraction, trend synthesis, cross-page comparison, unit conversion and multi-step arithmetic beyond OCR.
Video & long-horizon reasoning14%Event localization, temporal causality, audio-visual alignment, dense captions and frame consistency.
GUI agents & tool use14%Screenshot understanding, click-coordinate accuracy, cross-app long-horizon planning, recovery and multimodal tool orchestration.
Generation & editing6%Text-to-image, instruction editing, multi-turn consistency, Chinese glyph rendering and interleaved generation.
Chinese context & culture6%Chinese memes, heritage sites, policy documents, dialect speech and local everyday context.
Ecosystem, cost & availability8%API stability, open weights, on-prem and domestic-chip support, docs, rate limits and real unit price.

Evidence grades

Two 92s, two levels of confidence

A

Cross-verified

At least one independent board or reproduction agrees with official claims; the technical report discloses training and evaluation details.

14 models
B

Partially verified

Mainly official reports with limited community reproduction; fast version churn or incomplete public benchmarks.

16 models
C

Single source

Vendor self-test or a single media account with no reproducible material. Such models are down-weighted and marked “watch”.

2 models

Tiers

Score bands and tiers

Tiers are communication tools, not verdicts. Gaps under 0.5 count as level; crossing a tier usually means a structural gap in at least one dimension.

S+≥ 94Category leader; fit for unattended general pipelines
S92 – 93.9Tier 1; may already lead a single dimension
A+89 – 91.9Strong challenger with a price or ecosystem edge
A85 – 88.9Reliable workhorse with a manageable gap
A-82 – 84.9Specialist; narrow coverage
B+78 – 81.9Viable alternative; PoC required
B< 78Watchlist; not for direct production

Limitations

What this method cannot do

  • No closed-lab replication: scores map public evidence, not same-machine, same-condition runs.
  • Information lag: vendors iterate faster than public benchmarks.
  • Price is a snapshot: volume discounts, caching and batching are excluded.
  • No human preference: no pairwise blind tests, so style, tone and usability are not scored.
  • Chinese weight may still be low for global products: drop it to 0–2% when serving non-Chinese users.

Updates

Updates, corrections and citation

Cadence

Regular review on the first Monday of each month; major releases enter observation within 72 hours and become official after 30 days.

Corrections

Verifiable evidence is reviewed within seven days; the changelog keeps old and new scores. No silent edits.

Citation

CN Multimodal Tier List. (2026). 2026 China Native Multimodal LLM Tier List. 2026cnmllmtierlist.datacorp.top

About the method

Why these eight dimensions?

They cover the usual multimodal selection variables: what goes in (perception, documents, charts, video), what comes out (agents, generation), who it serves (Chinese) and how it ships (ecosystem). More dimensions dilute weights; fewer hide key differences.

Are scores human or model-generated?

Scores are set by the editorial desk against a fixed rubric, using public benchmarks, official reports and community reproductions. AI assists with research and drafts, never with the final score. Each model carries an evidence grade; A means third-party cross-verification.

Why do two 90-point models feel so different?

The composite is a weighted sum and excludes latency, stability, prompt sensitivity and product UX. Two models can tie overall while one leans documents and the other video. Read dimension scores and the scenario matrix.

Can I challenge a score?

Yes. Send reproducible evidence — a public benchmark, a reproduction script or failure cases — via the about page. We review within seven days and publish the outcome in the changelog.

Is this site sponsored by vendors?

No. We do not sell rankings or run affiliate promotions. Every model follows the same rules; any commercial placement is clearly labelled and never affects scores.

Personal ranking · Not an official score

Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.