Traceable changes
Changelog
Every weight change, score move and correction is logged. Old scores stay visible so you can check whether our judgement is consistent.
Content expansion: five new essays, 28 in total, 170 pages
- Five new essays — naming archaeology, the Chinese dialect tier list, context inflation, the self-knowledge audit and the anti-sycophancy index — plus the eight data-series essays not previously logged, taking the essay total from 15 to 28.
- Five new recomputable artefacts: the five-family naming table, the eight-class dialect board, the nominal-versus-effective context depth table, the self-knowledge bias matrix and the capitulation matrix.
- Pages rise from 144 to 170; models and score points are unchanged (32 models / 256 points) — this cycle is content expansion and does not touch the board data.
- Corrected the OG images and site tagline from “26 models / 208 score points” to “32 models / 256 score points”.
Second expansion: 32 models, 15 essays, 144 pages
- Six industry-scenario models added: Hikvision Guanlan 3 (surveillance), Cloudwalk Congrong 2 (recognition), Zidong Taichu 2.0 (academic all-modality), Ant Ling 2 (finance), JD Yanxi 2 (e-commerce) and NetEase Fuxi 2 (games); score points rise from 208 to 256.
- Three new essays — the surveillance-video field guide, the industry exclusivity map and the cross-review experiment — take the total from 12 to 15; pages rise from 126 to 144.
- The evidence spectrum becomes 14A / 16B / 2C; open-weight models rise from 11 to 12.
- Pre-registered falsification conditions are adjudicated on the 144-page basis; the expansion predates the first Search Console crawl.
Day-one expansion: 26 models, 12 essays, 126 pages
- Sample expanded before first crawl: added InternVL 4.0 and NeuroXess-N1 as cold-model entries; score points rise from 192 to 208.
- Two new essays — the frontier wave and why cold models survive — take the total from 10 to 12; pages rise from 118 to 126.
- The evidence spectrum moves from 13A/11B/0C to 14A/11B/1C: the site’s first grade-C model makes the evidence column discriminating.
- Pre-registered falsification conditions are adjudicated on the 126-page basis; the expansion predates the first Search Console crawl, so the goalposts did not move.
Site launch, cold-start experiment begins
- Published 24 Chinese multimodal models, 8 dimensions and 192 score points.
- Shipped 10 essays, 4 tools and 10 comparison pages: 118 static pages across Chinese and English.
- Submitted sitemap, llms.txt and structured data; the experiment log starts here.
Four frontier models arrive; the board reaches 24
- Added Qwen3.8-Max, Qwen3.8-Flash, DeepSeek-V4.1-Flash and GLM-5.3 — all released between late September and early October 2026.
- All four are evidence grade B: vendor-reported benchmarks, no independent reproduction, under 30 days of observation. Score points rise from 160 to 192.
- Qwen3.8-Max takes #1 at 95.4 and Doubao Seed 2.0 Pro moves to #2 at 95.2 — inside our 0.5-point same-band threshold, so first place is a tie.
- Comparison pages grow from 6 to 10, adding Qwen3.8-Max vs Doubao, GLM-5.3 vs GLM-5 and DeepSeek-V4.1-Flash vs Qwen3.8-Flash among others.
Data freeze and final weights
- Evidence baseline: SuperCLUE-VLM April and October 2026, official reports and community reproductions.
- Final weights: perception 20 / documents 18 / charts 14 / video 14 / agents 14 / generation 6 / Chinese 6 / ecosystem 8.
- Every model graded A/B/C; 13 reach grade A.
Expanded from 12 to 20 models
- Added Meituan LongCat-Next, BAAI Emu3.5, AgiBot WITA-Omni, Huawei Pangu 5.5, MiniMax M2.5, ModelBest MiniCPM-V 4.5, BAGEL-2 and StepFun Step-3.
- Added document specialists, embodied, edge and open unified models so the list is not all general flagships.
Rubric revision: agent weight increased
- Agent weight raised from 10% to 14%; generation cut from 10% to 6%. Rationale: in enterprise procurement, GUI/tool use is growing faster than image generation.
- Impact: Kimi K3 +0.3, Doubao +0.2, Emu3.5 −0.2.
Visualisation system completed
- Shipped radar, heatmap, scatter, trend and delta charts — all server-rendered SVG with no chart library.
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.