Three things to remember
- Divergence comes from three things: modality coverage, the evidence discount, and leaderboard absence. Public boards test the questions a model can answer; a tier list tests whether it can do your job.
- Specialists diverge most. DeepSeek-OCR 2 is #1 on documents alone and #20 overall — a 19-place gap that is entirely a coverage penalty.
- InternVL 4.0 shows a new kind of divergence: a permanent top-three fixture in open-source boards that simply does not exist in Chinese comparisons. Its perception rank is 4th and its video rank 26th — a 22-place internal spread inside one model.
There are already too many essays about how leaderboards mislead: data contamination, prompt differences, sample leakage — every one lists seven or eight. The problem with that genre is that it tells you leaderboards are untrustworthy without telling you how untrustworthy, and for which models.
This essay takes a different approach: quantify the gap. For all 32 models on the 2026 China native multimodal LLM tier list we compute two divergence indices — one measuring disagreement with public benchmarks, one measuring disagreement with itself. The second is the main artefact here, and it is something public boards never publish.
Three kinds of divergence
External divergence: the distance between a model’s rank on public boards (SuperCLUE-VLM, OmniDocBench and others) and its rank here.
Internal divergence: the distance between a model’s rank on its strongest dimension and its overall rank.
Absence divergence: the gap between a model’s presence in public benchmarks and open-source community boards and its presence in Chinese comparison pieces. This class only became worth separating in 2026 — the open-source community runs its own popularity boards, the comparison circle runs its own fixed shortlists, and the two barely overlap.
External divergence measures “how much two framings disagree”; internal divergence measures “how much this model depends on you looking at one dimension”. The second is more useful for selection because it is fully recomputable and covers all 32 models.
Internal divergence: rank gaps from 0 to 27
Definition: gap = overall rank − strongest-dimension rank. The larger the positive number, the more “strong in one dimension, weak in aggregate” the model is.
| Model | Overall | Strongest dimension | Gap | Dimension rank spread |
|---|---|---|---|---|
| BAGEL-2 | 31 | Generation #4 (90) | +27 | 28 |
| Emu3.5 | 24 | Generation #1 (98) | +23 | 27 |
| Hikvision Guanlan 3 | 27 | Video #7 (91) | +20 | 25 |
| Cloudwalk Congrong 2 | 32 | Perception #12 (91) | +20 | 18 |
| DeepSeek-OCR 2 | 20 | Document #1 (99) | +19 | 30 |
| WITA-Omni Preview | 23 | Video #6 (92) | +17 | 26 |
| Ant Ling 2 | 22 | Chinese #9 (92) | +13 | 22 |
| NetEase Fuxi 2 | 29 | Generation #16 (82) | +13 | 16 |
| Keye-VL-2.0 | 14 | Video #2 (94) | +12 | 20 |
| InternVL 4.0 | 16 | Perception #4 (96) | +12 | 26 |
| JD Yanxi 2 | 28 | Chinese #16 (89) | +12 | 16 |
| NeuroXess-N1 | 30 | Chinese #19 (88) | +11 | 12 |
| Qwen3.8-Flash | 13 | Ecosystem #3 (97) | +10 | 18 |
| MiniCPM-V 4.5 | 26 | Ecosystem #17 (80) | +9 | 15 |
| MiniMax M2.5 | 19 | Agent #11 (86) | +8 | 18 |
| Zidong Taichu 2.0 | 25 | Video #17 (84) | +8 | 13 |
| LongCat-Next | 9 | Generation #2 (92) | +7 | 22 |
| Spark X2 | 18 | Chinese #12 (90) | +6 | 12 |
| DeepSeek-V4.1-Flash | 12 | Document #6 (95) | +6 | 19 |
| Step-3 | 17 | Agent #12 (85) | +5 | 14 |
| Kimi K3 | 5 | Agent #1 (95) | +4 | 11 |
| GLM-5 | 6 | Chinese #2 (95) | +4 | 9 |
| MiMo-V2.5 | 7 | Video #3 (93) | +4 | 11 |
| ERNIE 5.1 | 11 | Chinese #7 (93) | +4 | 9 |
| SenseNova 6.5 Pro | 8 | Perception #6 (94) | +2 | 8 |
| Qwen3-Omni-30B | 10 | Generation #8 (88) | +2 | 6 |
| Huawei Pangu 5.5 | 21 | Document #19 (87) | +2 | 9 |
| Doubao Seed 2.0 Pro | 2 | Perception #1 (97) | +1 | 6 |
| Qwen3.5-397B | 3 | Document #2 (97) | +1 | 3 |
| Hunyuan-Vision | 15 | Generation #14 (82) | +1 | 5 |
| GLM-5.3 | 4 | Agent #3 (95) | +1 | 9 |
| Qwen3.8-Max | 1 | Chart #1 (96) | 0 | 4 |
Five things read directly off this table:
Still only one model has a gap of 0. Qwen3.8-Max’s strongest-dimension rank (chart #1) exactly matches its overall rank (#1). It is the only model of 32 that is top-tier on a single dimension and first overall. That is the hardest argument for its number-one spot — more persuasive than the number 95.4 itself.
Models with gaps above 15 grew from four to six. BAGEL-2 (generation #4, 31st overall), Emu3.5 (generation #1, 24th), Guanlan (video #7, 27th), Congrong (perception #12, 32nd), DeepSeek-OCR 2 (document #1, 20th) and WITA-Omni (video #6, 23rd). If your job is image generation, document parsing, surveillance video or recognition, choosing off the overall board excludes the strongest options outright.
The largest spread is DeepSeek-OCR 2 at 30 places — document #1 to generation #31. A single model’s internal asymmetry can reach that: an undisputed champion on document benchmarks, 20th on a board that rewards doing a bit of everything.
The two industry models are new specimens of absence divergence. Guanlan is 7th on video and 27th overall: a workhorse in security industry solutions, absent from public benchmarks because surveillance data cannot be published. Congrong is 12th on perception and 32nd overall: steady in face-recognition acceptance tests, last in comparisons. For both, the general board’s rank carries almost no weight in the procurement decision.
InternVL 4.0 and Ant Ling 2 represent two other divergences. The first is the inversion of community heat and comparison presence (perception #4, video #30, spread 26); the second is scenario corpora against general capability (Chinese #9, video #31, spread 22) — its Chinese 92 is third on the site, yet it ranks 22nd overall.
Absence divergence: the InternVL 4.0 phenomenon
External divergence can only be computed for models present in both framings, which exposes the biggest problem with public boards: cross-version scores are incomparable — and, more subtly, comparison samples are homogeneous.
The SuperCLUE-VLM scores we cite come from two different cycles:
| Model | SuperCLUE-VLM | Cycle | Our composite | Our rank |
|---|---|---|---|---|
| Qwen3.8-Max | 91.8 | 2026-10 | 95.4 | 1 |
| GLM-5.3 | 89.4 | 2026-10 | 92.4 | 4 |
| Qwen3.8-Flash | 84.2 | 2026-10 | 87.5 | 13 |
| Doubao Seed 2.0 Pro | 90.66 | 2026-04 | 95.2 | 2 |
| Qwen3.5-397B | 88.50 | 2026-04 | 94.6 | 3 |
| SenseNova 6.5 Pro | 85.88 | 2026-04 | 89.9 | 8 |
Look at the cycle column. Sorting the 2026-04 and 2026-10 scores in one table yields “Doubao 90.66 > GLM-5.3 89.4” — but those two numbers are six months, one full benchmark revision and one GLM architecture upgrade apart. That sort is invalid, and it is exactly what most aggregator boards do.
Only same-cycle comparison is meaningful: within the 2026-10 cycle, Qwen3.8-Max 91.8 > GLM-5.3 89.4 > Qwen3.8-Flash 84.2, directionally consistent with our ranks (1, 4, 13) but with absolute gaps compressed. Public boards’ score distributions are flatter than real capability distributions — an inherent property of multiple-choice benchmarks, which averages away a strong model’s long-tail failures.
Absence divergence is a different animal. InternVL 4.0 has been top-three on Hugging Face downloads among Chinese multimodal models for a long time and first on OpenCompass’s open-weights board, yet it barely exists in Chinese comparisons — because the standard comparison move is “pull ten flagship names”, and open-source entries are not on that list. SenseNova, Step-3 and NeuroXess-N1 are absent for different reasons (no general marketing, undisclosed architecture, no API), but the result is identical: the capability distribution you can see on public boards is a biased sample of the real one.
The cost of biased sampling is paid by the buyer: a team on an open-source stack that selects off comparisons alone will miss the open-source perception ceiling entirely.
Coverage penalty: where DeepSeek-OCR 2’s 19 places went
This is the largest internal divergence case and deserves its own section.
DeepSeek-OCR 2’s document score of 99 is the highest on the site, OmniDocBench v1.5 reaches 91.09%, and the output price is ¥0.8 per million tokens. On “document parsing” alone it is the undisputed first.
But its modalities are text and image only: no video, no audio, no agents, and generation close to zero. Under our eight-dimension weights, the combined 40 points of video 14, agents 14, generation 6 and Chinese 6 yield very little, dragging the composite to 81.9 and overall rank 20.
Those 19 places are not it getting weaker — they are it being punished for things it does not do.
So the correct usage is as a component, not a candidate: let it do first-pass parsing in the document pipeline and have a flagship review. The DeepSeek-V4.1-Flash versus DeepSeek-OCR 2 comparison breaks down the cost structure of that pairing.
Evidence discount: reverse divergence for fresh models
Fresh frontier models show a special kind of external divergence: vendor-benchmark rank far above the certainty implied by the tier-list rank.
Qwen3.8-Max, GLM-5.3, DeepSeek-V4.1-Flash and Qwen3.8-Flash all shipped between late September and early October 2026, and all four are evidence grade B rather than A here. Two things are missing: independent third-party reproduction, and 30-plus days of stability observation.
That produces a seemingly contradictory result: Qwen3.8-Max is first on SuperCLUE-VLM 2026-10 and first here, while this site simultaneously tells you “this number one is less certain than the runner-up Doubao”.
That is not caution — it is pricing two risks separately. Capability risk is carried by the composite; certainty risk by the evidence grade. Experience says this reverse divergence usually converges after one evaluation cycle (about three months): either third-party reproduction confirms and the grade rises to A, or version differences surface and the score drops. Every grade change is logged in the changelog.
How to select with the divergence index
Three steps:
- Set your own tolerance for internal divergence. If your workload touches one or two dimensions, tolerance can be high — sort by the single dimension and ignore the overall board. If it spans four or five, tolerance must be low and you look only at models with a spread of 5 or less (Qwen3.8-Max, Qwen3.5-397B, Doubao, GLM-5.3, Hunyuan-Vision).
- Check that public-board cycles match. If they do not, do not compare across; compare within a cycle only.
- Give fresh models one evaluation cycle of observation. If the project timeline allows, wait for third-party reproduction; if it does not, take the grade-A runner-up and buy back the certainty.
Dimension sorting, spreads and evidence grades are all switchable on the tier list page. The full divergence computation ships with the site data; its input is the 256 dimension scores of the 32 models.
FAQ about this essay
What do positive and negative divergence mean?
Positive means it ranks better on public boards than here — overrated by public benchmarks. Negative means it ranks better here — underrated by them. The larger the absolute value, the more the two framings disagree and the more you need your own scenario weights.
Why not just adopt one public leaderboard’s ranking?
Public boards aim to compare capability ceilings; selection aims to match a scenario. The former never needs your image resolution, concurrency or compliance constraints — and without them the latter is meaningless. We treat public scores as evidence inputs, not conclusions.
