Lead essay
A new #1 — and eight “no second place” models at the bottom of the board
30-second answer
The top of the board changed this cycle. Qwen3.8-Max scores 95.4, Doubao 95.2 — a gap of 0.2. Our method treats anything under 0.5 as the same band, so this is not an overtake, it is a tie. Perturbing all eight dimension weights by ±30% across 2000 runs puts Qwen first in 91.1% of reasonable weight vectors and Doubao first in 8.9%.
What separates them is not the score but the certainty. Qwen3.8-Max is grade B: vendor-reported benchmarks, under 30 days old, no independent third-party reproduction. Doubao is grade A. That is why this site prices capability risk and certainty risk separately — the composite tells you who is stronger, the grade tells you how likely "stronger" survives the next release. For a three-year contract, trading 0.2 points for a grade-A trail is usually a good deal.
Two expansions: from technology-form cold to industry-scenario cold
The first expansion improved technology-form coverage: InternVL 4.0 (85.0, #16) is a perception specialist — top-three downloads in the open-source community, nearly absent from Chinese comparisons, perception 96, MIT, grade A; NeuroXess-N1 (77.2, #30) puts EEG/fNIRS signals, images and text in one token space — the only grade-C model here and the only entrant in neural-signal decoding. The second expansion supplies the other half, industry scenarios: Hikvision Guanlan (surveillance, grade C), Cloudwalk Congrong (recognition), Zidong Taichu (academic all-modality), Ant Ling 2 (finance, Chinese 92), JD Yanxi 2 (e-commerce images) and NetEase Fuxi 2 (game content). All six composites sit between 76 and 81, yet each is a “no second place” in its scenario. The evidence spectrum reads 14A / 16B / 2C — the evidence column discriminates for the first time.
The most commercially interesting of the four frontier releases is DeepSeek-V4.1-Flash: ¥0.6 per million output tokens, document 95, open weights, and the second-cheapest point on the 1024px cost-per-1,000-images curve at ¥0.54. The only model below it (MiniCPM-V 4.5 at ¥0.45) scores 79.6 overall. This is the first time "cheap" and "capable" land on the same model. Its limits are equally clear — video 78, generation 72 — so do not use it for video QA or creative work.
Thirteen months of the leaders
Editorial mappings, re-scored under one weighting; not launch-day vendor numbers.
← scroll for the full trend →
| Top five | Overall | Docs | Video | ¥/M |
|---|---|---|---|---|
| Qwen3.8-Max S+ | 95.4 | 97 | 93 | ¥18 |
| Doubao Seed 2.0 Pro S+ | 95.2 | 96 | 96 | ¥15 |
| Qwen3.5-397B-A17B S+ | 94.6 | 97 | 92 | ¥12 |
| GLM-5.3 S | 92.4 | 94 | 88 | ¥9 |
| Kimi K3 S | 92.3 | 92 | 90 | ¥16 |
Why the document specialist ranks 20th
DeepSeek-OCR 2 scores 81.9 overall and ranks 20th, yet its document dimension is 99 — the highest in the field, two points above Qwen3.8-Max. The reason is coverage: no video, audio, generation or agents, so the weighted score drags it down.
That is both the value and the trap of a weighted list. If documents are your only workload, raise document weight from 18% to 60% and zero out video, generation and agents. DeepSeek-OCR 2 enters the top three; the overall champion may leave the top five. A ranking is a set of editable assumptions. For specialists, the correct move is to read the dimension sort directly — the overall rank has near-zero predictive power for them.
How to use it
- Sort by your primary scenario: documents, video or agents.
- Read dimension scores for the top three, not the composite.
- When the scenario is cold (industrial QC, vertical classics, neural signals, open-source deployment), check the mini tier list and the cold-model matrix before the overall board.
- Add deployment, budget and Chinese constraints in the picker; the shortlist usually falls to 3–5.
- Run 50–100 real samples in a PoC; if it conflicts with the list, trust the PoC.
We deliberately avoid a single correct answer. Change the weights and the winner changes — this cycle's 91.1% versus 8.9% is the proof. Not because the list is unstable, but because selection should work that way.
