Chinese multimodal models and this ranking, explained

On our eight-dimension weighting, Qwen3.8-Max leads at 95.4 (evidence B) and Doubao Seed 2.0 Pro is second at 95.2 (evidence A), followed by Qwen3.5-397B-A17B (94.6), GLM-5.3 (92.4) and Kimi K3 (92.3). Gaps under 0.5 count as one band, so the top two are tied. Single-dimension leaders differ: DeepSeek-OCR 2 for documents, Doubao for video, DeepSeek-V4.1-Flash for value.

The list focuses on models built by Chinese teams and legally available in mainland China. Global models appear in separate comparison essays, not mixed into this ranking.

Modality coverage and long-task stability. S+ has no visible hole across eight dimensions and suits unattended general pipelines; S is usually one or two dimensions weaker and needs scenario validation.

Structured desk evaluation from public evidence, not a closed-lab test. Humans set scores against a rubric, with A/B/C evidence grades. We do not claim tests we did not run.

The composite covers eight dimensions. It has no video, audio, generation or agent support, so coverage drags the total. For documents only, 99 and ¥0.8/M is exceptional value.

Yes. All weights are published, the picker re-weights by scenario, and you can copy the dataset and compute your own.

Qwen3.8-Max for overall quality, Doubao Seed 2.0 Pro if you need a grade-A trail; Kimi K3 or GLM-5 for on-prem; DeepSeek-OCR 2 for document batches; Doubao for video; DeepSeek-V4.1-Flash for budget-sensitive work.

Depends on the licence. Apache-2.0 is the most permissive; community and research licences often restrict commercial use or require registration. Check the original text with legal before shipping.

Reviewed monthly, and within 72 hours of a vendor price change where possible. We use public list prices and exclude volume discounts and credits.

Yes, under CC BY 4.0. Credit “CN Multimodal Tier List” and link to the source page. Contact us for bulk or commercial reuse.

Yes. The experiment page publishes the hypothesis, build setup, submission dates and metrics. We will not fabricate a “#1 in seven days” result; failures will be logged too.

No guarantee. With Search Console verification, a complete sitemap and internal links, days to weeks is common — but sometimes only the homepage is indexed. The experiment page tracks what actually happens.