Three things to remember
- Of the 15 scenarios, six have exactly one contender: surveillance (Guanlan), neural signals (NeuroXess-N1), academic all-modality (Zidong Taichu), e-commerce product images (Yanxi 2), game content (Fuxi 2) and financial instruments (Ling 2, whose only rivals are flagships).
- Scenario exclusivity correlates near-zero with overall rank: the 15 exclusive models average 24th on the overall board, yet every scenario they cover has real paying demand.
- The map is meant to be read backwards: find your scenario on the map first, then return to the tier list for that scenario’s one or two entries — selecting off the overall board systematically misses the only answer.
Comparisons ask “who is stronger”; procurement asks “who can do my scenario”
This 2026 China native multimodal LLM tier list has an unused reading all along: re-sort by scenario, not by score.
The reason is a simple observation: at least 15 of the 32 models appear in no mainstream Chinese comparison. They spread across a dozen-plus vertical scenarios, with only one or two contenders each. The comparison ranking has almost no predictive power for them — 27th overall is first on the surveillance mini board, 32nd overall is mid-price first in face recognition, and 30th overall is the only neural-signal decoder anywhere.
This essay is that map: 15 scenarios, each with the models that can compete, evidence grades, procurement entries, and their overall-board ranks. The reading is reversed — find your scenario on the map first, then return to the tier list for that one or two.
The map method: three conditions
A model enters a scenario’s “can compete” list only if it satisfies all three:
- Verifiable scenario accumulation — public vertical cases, industry delivery records or checkable corpus building, not a launch-deck line claiming “optimised for some industry”;
- A top-six score on the corresponding dimension here — scores are public on the dossiers and recomputable;
- No third model satisfying both — that is the definition of “exclusive”.
The standard is fixed here. You may disagree, but you can recompute. Evidence criteria are on the methodology page.
The industry exclusivity map
| # | Scenario | Models that can compete | Evidence | Overall | Procurement entry |
|---|---|---|---|---|---|
| 1 | Government / domestic chips | GLM-5.3, GLM-5 | B / A | #4 / #6 | Closed + on-prem licence (Ascend images) |
| 2 | Financial instruments / compliance | Ant Ling 2 | B | #22 | Closed commercial API |
| 3 | E-commerce images / supply chain | JD Yanxi 2 | B | #28 | Closed commercial API |
| 4 | Games / content production | NetEase Fuxi 2 | B | #29 | Closed commercial API |
| 5 | Surveillance | Hikvision Guanlan 3, Cloudwalk Congrong 2 | C / B | #27 / #32 | Industry licensing |
| 6 | Medical imaging assistance | SenseNova 6.5 Pro, Huawei Pangu 5.5 | A / B | #8 / #21 | Closed commercial + industry solutions |
| 7 | Industrial QC | SenseNova 6.5 Pro | A | #8 | Closed commercial + industry solutions |
| 8 | Chinese documents / OCR | DeepSeek-OCR 2 | A | #20 | Open source (MIT-family) + API |
| 9 | Long-video timing | Doubao Seed 2.0 Pro, WITA-Omni | A / A | #2 / #23 | Closed commercial / filed |
| 10 | GUI agents | GLM-5.3, Qwen3.8-Max, Kimi K3 | B / B / A | #4 / #1 / #5 | Closed commercial / open weights |
| 11 | Open-source perception | InternVL 4.0 | A | #16 | MIT weights + API |
| 12 | Edge at low cost | MiniCPM-V 4.5 | A | #26 | Apache-2.0 (registration) |
| 13 | Embodied / native-grade research | WITA-Omni | A | #23 | Closed, filed |
| 14 | Neural-signal decoding | NeuroXess-N1 | C | #30 | Project licensing |
| 15 | Academic all-modality research | Zidong Taichu 2.0 | B | #25 | Research licence |
How to read it, row by row:
Row 1 is the only “two-model scenario”. GLM-5.3 and GLM-5 both satisfy all three conditions for government and domestic chips — chips, compliance, capability. The only double slot on the map, because Zhipu alone holds both the full Ascend stack and the most public-sector deliveries; everyone else holds one.
Rows 2–4 are this cycle’s new scenario exclusives. Ling 2 (Chinese 92 + finance corpora), Yanxi 2 (product-image + SKU corpora), Fuxi 2 (generation 82 + video 84) — all three rank 22nd or lower and are mid-field on general dimensions, yet each scenario has no second model satisfying condition three. They prove scenario corpora can form exclusivity, and exclusivity does not depend on general scores.
Row 5 is the only “hardware-vendor scenario”. Neither Guanlan nor Congrong enters the market through an API; the procurement entry is industry licensing — which makes their procurement flow completely different from API models (see G2 of the veto-gate tier list).
Rows 6–7 are the last stand of “professional vision”. SenseTime satisfies both medical assistance and industrial QC on a decade of vertical cases — precisely the part public benchmarks can never reach (the data cannot be public).
Rows 8–10 are the boundary between “true” and “false” exclusivity. DeepSeek-OCR 2’s document 99 is first on the site (true exclusivity); long-video timing has two contenders; GUI agents have a three-way tie at 95 — yet by failure-recovery rate only GLM-5.3 at 78% is recommended (see the GUI agent benchmark). “Several models can compete” and “only one should be chosen” are different questions; the map answers only the first.
Rows 11–15 are the long tail within the long tail. Open-source perception, edge, embodied, neural signals, academic — one model each, all with explicit procurement entries (MIT, Apache, filed, project, research). These five rows share one property: their traffic is near zero, but for readers who need them, they are the only answer on the web.
A counter-intuitive observation
Fifteen scenarios, eighteen exclusive slots — these models average 24th on the overall board. Yet every scenario they cover has real paying demand.
This is not the ranking failing; it is two different products: comparisons sell “capability ceiling” and sort by it; the scenario map sells “irreplaceability” and sorts by that. Both products have buyers, but the buying differs completely — using a capability-ceiling ranking for scenario procurement systematically misses the only answer; using the scenario map to compare capability ceilings buys a pile of models that are good at everything and best at nothing.
How to use this map
Three steps:
- Find your primary scenario on the map. Not there? You are probably on the general battlefield — return to the tier list and sort by dimension.
- Check that row’s evidence grade and procurement entry. Grade C (Guanlan, NeuroXess-N1) means every figure is vendor-reported and acceptance must run on your own site data; industry licensing means solution procurement, not API procurement.
- Return to the tier list and check that one or two models’ dimension scores. The map tells you who can compete in the scenario; dimension scores tell you whether it drags on the other dimensions your business touches — the latter decides whether you pay a generality price for scenario exclusivity.
If you need two scenarios (say finance plus long video), the map’s answer is a combination, not a single choice: Ling 2 for instruments, Doubao for video, with a structuring layer between. The cost of combination is integration; the benefit is using the exclusive model in every scenario — the standard shape of multimodal procurement in 2026.
Boundaries and method
The three “can compete” conditions are stated at the top and recomputable: dimension scores are public on the dossiers, and the basis for scenario-accumulation judgements is in the per-model records of why cold models survive. The map updates with versions — if a third model satisfies a scenario’s conditions, or an exclusive scenario gets covered by general capability, we revise the map and log the reason in the changelog.
FAQ about this essay
How is this map different from “why cold models survive”?
The perspective is reversed. “Why cold models survive” is model-first: model by model, why it is cold and how it survives. This map is scenario-first: from your industry, asking “who can do this scenario”. The first answers “why is it on the board”, the second “who should I look at” — the same data, two reading directions.
What is the “can compete” standard on the map?
Three conditions together: verifiable delivery or corpus accumulation in the scenario, a top-six score on the corresponding dimension here, and no third model satisfying both. The standard is fixed on the page — you may disagree, but you can recompute it: all dimension scores are public on the dossiers.
