Three things to remember
- Cold models share three causes: no general marketing, low participation in public benchmarks, and project-based rather than API-based procurement. All three point the same way — their heat is not in the search box, it is on the delivery site.
- The irreplaceability matrix shows each of the four deep-dives owns a scenario with no runner-up: NeuroXess-N1’s neural-signal decoding is the board’s only entry; InternVL 4.0’s perception 96 is the open-source ceiling; SenseTime’s industrial QC and Step-3’s four-modality API surface are the last options at mid price.
- The procurement implication is direct: cold models missing your POC is normal, but when the scenario matches, they are the only answer. The mini tier list and the veto gates are the two tools built for “only answer” scenarios.
Every public comparison tests the same dozen models
Stack the participant lists of every “Chinese multimodal LLM comparison” on the 2026 Chinese internet and the overlap is striking: Doubao, Qwen, Kimi, GLM, DeepSeek, ERNIE, Hunyuan, MiniMax — eight to ten names, on rotation. Not because the authors are lazy, but because rankings are content products and run on ROI: testing a model with no public API is useless to readers; testing an edge-only model is useless to traffic.
The 2026 China native multimodal LLM tier list on this site covers 32 models. Of the extra entries, at least fifteen appear in no mainstream Chinese comparison: SenseTime, StepFun, NeuroXess, InternVL, Kuaishou, Meituan, AgiBot, Huawei and ModelBest.
This essay does not rehabilitate them. Some genuinely do not rank high, and two have no public API at all. It answers one question: when a model holds no popular slot on any board, what earns it a test and a purchase order.
The answer in preview: having no second place.
Four shared causes of coldness
Go through the fifteen cold models’ dossiers and the causes of absence are classifiable:
| Cause | Models | Mechanism |
|---|---|---|
| No general marketing | SenseNova, Step-3, Cloudwalk Congrong | Launch events target government and industry customers, not developer communities, so public volume is structurally low |
| Procurement entry is not an API | NeuroXess, Huawei Pangu, Hikvision Guanlan | Project or industry licensing; no standardised call entry, so comparisons cannot plug-and-test |
| Poor fit for leaderboard sampling | InternVL, Keye-VL, LongCat-Next, WITA-Omni, MiniCPM, Zidong Taichu | Open-source entries, video specialists, embodied, edge, academic — none in the “ten flagship names” sampling frame |
| Scenario binding too deep | Ant Ling 2, JD Yanxi 2, NetEase Fuxi 2 | Scenario corpora dominate training data, so they lose on general boards by construction — and the vendor has no incentive to break out |
All four causes point the same way: these models’ heat is not in the search box, it is on the delivery site. And the search box is the shared eye of comparison authors and readers. Sample through the search box and you get not a market overview but a map of marketing budgets.
SenseNova 6.5 Pro: vertical cases are the oldest moat
8th overall, 89.9, grade-A evidence, ¥12 — a respectable record by any comparison’s standard, yet SenseTime appears in none of them. The reason is plain: SenseNova’s launch audience is tertiary hospitals and automotive lines, not developers.
Its position in the Chinese document tier list nobody benchmarks is its real portrait: first on tongue diagnosis, first on industrial QC, second on opera masks. What those three classes share is expert annotation, unpublishable data, and no overseas demand — exactly where public benchmarks never reach. SenseTime’s decade of vision accumulation happens to sit there.
Its weakness is equally honest: on the Pareto frontier’s 9-dimension strict dominance table it is dominated by same-priced Qwen3.5-397B — it genuinely loses the general battle, leading on none of the eight dimensions. Its entire survival case lies outside the general battlefield.
Procurement implication: if your multimodal needs touch medical imaging assistance or industrial QC, SenseNova belongs on the shortlist even though it appears in almost no comparison — that is sampling bias, not capability absence. Conversely, for general QA plus chart understanding, same-price strict optima exist and the “SenseTime” name is not worth the money.
StepFun Step-3: the fullest mid-price interface, the blackest architecture box
17th overall, 84.7, grade B, ¥5. Step-3’s absence rate from Chinese comparisons is nearly 100%, for the opposite reason from SenseTime: it has no shortage of volume — it lacks evaluability. Four-modality interfaces are listed in full; architecture details are not public; native grade can only be provisioned N2 by observable behaviour.
Put it back into its price band, though, and an awkward fact emerges: it alone holds the ¥5 tier. Cheaper options (Qwen3.8-Flash ¥1.2, DeepSeek-V4.1-Flash ¥0.6) have stronger ecosystems but thinner verticals; dearer options (GLM-5.3 ¥9) are a tier stronger but also a tier harder to procure. At the intersection of “four-modality interface + mid price + closed commercial”, it is the only product among the 26 here.
Its survival mode is classified in the cost essay as: no strictly leading dimension; survives on ties alone — the most fragile survival mode, since any vendor’s next version can erase it. But “most fragile” is not “currently worthless”: a project needing audio + video + image on one interface without flagship pricing has exactly one candidate today.
Procurement implication: treat Step-3 as “interface insurance” — when the modality combination you need has no other mid-price product, it is the only option; when Qwen or GLM fills that gap at the same price, it leaves the shortlist. Choosing it means accepting the integration risk of an architecture black box and locking interface-compatibility terms in the contract.
NeuroXess-N1: the only neural-signal modality on the web, and one of two grade-C models
25th overall, 77.2, grade-C evidence, no public API. This is the most “off-centre” entry among the 26 here, and the one whose inclusion is most disputed.
Facts first: an implantable-BCI company put EEG/fNIRS temporal signals, images and text into one discrete token space for neuro-rehabilitation assessment and embodied interaction. Its scenario score of 88 for neural-signal decoding is the only entrant on the site — it did not win anything; nobody else showed up.
It is also one of two grade-C models here (the other is Hikvision Guanlan 3): every figure vendor-reported, zero third-party reproduction. By this site’s own standards, such a model should not be listed. It is, for reasons unrelated to score:
- Coverage is the value. When a modality has exactly one tested model, that fact is itself unique information on the web. Homogeneity in comparisons is killing exactly this kind of information.
- The C label is a filter, not a fig leaf. We publish evidence grade independently of the composite precisely so readers can decide for themselves whether to touch an all-vendor-figure model. Hiding it and promoting it to A are the same dishonesty.
- It stress-tests the board. A ranking that lists a grade-C BCI-company model at 25th while labelling it honestly is more auditable than one that covers only 14 grade-A/B flagships — the latter’s problem is sampling bias; the former’s is one extra row.
Its weaknesses must be stated plainly: ecosystem 62 is the lowest here, no SDK documentation, no rate-limit statement, generation 68, video 70 — outside neural and embodied scenarios it is unusable. Beyond neuro-rehabilitation, stroke follow-up and embodied interaction research, no business should consider it.
Procurement implication: this is the standard shape of an “only answer” scenario — the selection question is not “which model” but “whether to use it, and under what terms”. Under project licensing, a medical-grade data-compliance path is the real selling point; weights never leaving for the client is the real boundary.
InternVL 4.0: the inversion of community heat and ranking heat
16th overall, 85.0, grade-A evidence, MIT, ¥1 per million output tokens, perception 96 — the highest open-weights score below the closed flagships.
The most typical of the fifteen cold models: a permanent top-three download in the open-source community, with roughly zero presence in Chinese comparisons. We named the phenomenon in the rank-divergence index: absence divergence. The standard comparison move is “pull ten flagship names”, and flagship names are a function of marketing volume, not capability.
InternVL 4.0’s dossier is a complete specimen of that inversion: perception 96 (third on the site, behind two 97s), document 93, chart 90 — but video 66, third from the bottom, and agent 78 below the enterprise floor. Its capability profile is a sharp specialist, and specialists lose under any weighted overall board, while in the real procurement identity of “perception component” it is the open camp’s only option.
Another fact unique to it: in the 9-dimension strict dominance analysis it is an escaper — 2D-dominated by DeepSeek-V4.1-Flash (cheaper and 2.9 points stronger overall), yet strictly ahead of every cheaper rival on perception, 96 against 90. Translated into procurement: “the cheaper one has a higher composite” fails for open-source users, because an open-source buyer’s utility function puts licence type and self-hosting near 40% — and neither lives in the eight dimensions.
Procurement implication: open-source stack + perception component under ¥1 + MIT licence — when all three hold, there is exactly one candidate. It is wrong as a multimodal workhorse (video and agents are the priced-in trade), but as a component it is grade-A evidence plus the most permissive licence plus the open-source perception ceiling.
This cycle’s six industry models: a new kind of cold
The first four deep-dives are cold because of no marketing, no API, or poor leaderboard fit. The six new entries supply a fourth cause: scenario binding too deep. Their launch events address industry customers only, they lose on general boards by construction, and each holds a scenario with no second place.
| Model | Overall | Evidence | Composite / price | Home ground |
|---|---|---|---|---|
| Hikvision Guanlan 3 | #27 | C | 78.2 / industry licensing | Surveillance video: night-vision IR, perimeter, cross-camera tracking |
| Cloudwalk Congrong 2 | #32 | B | 76.1 / ¥7 | Recognition: faces, ID documents, receipts |
| Zidong Taichu 2.0 | #25 | B | 80.1 / no API | Academic all-modality: papers and ablations |
| Ant Ling 2 | #22 | B | 80.8 / ¥5 | Finance: instruments, compliance, risk documents |
| JD Yanxi 2 | #28 | B | 78.0 / ¥6 | Retail: product images, SKUs, supply-chain documents |
| NetEase Fuxi 2 | #29 | B | 77.5 / ¥4 | Content: game NPCs, fan creation, short video |
All six composites sit between 76 and 81 — precisely the band that looks like “should not be selected” on the overall board. But put each back in its home ground and the conclusion inverts: Guanlan has no second entrant on surveillance-video understanding; Congrong’s perception 91 is something its score peers cannot offer; Ling 2’s Chinese 92 trails only three flagships; Yanxi 2’s product-image corpora are exclusive among the 32; Fuxi 2’s generation 82 is the cheapest production option for content work; and Zidong Taichu is the only model outside the open-source community that publishes all-modality ablations.
They share something new as well: all six die at G2 or G3 of the veto-gate tier list — Guanlan at weight custody (industry licensing); Ling 2, Yanxi 2, Fuxi 2 and Congrong 2 at the capability floor (agent below 84); Zidong Taichu at commercial licensing (research licence). Cold models are not only hard for comparisons to see — they are hard for procurement processes to accept, and both usually trace to the same cause.
The irreplaceability matrix: fifteen cold models, fifteen “no second places”
Compressing the fifteen cold models into one matrix. The last column is this essay’s real thesis: if it disappears, which scenario has no fallback.
| Model | Overall | Evidence | Exclusive scenario | Procurement entry | If it disappears |
|---|---|---|---|---|---|
| SenseNova 6.5 Pro | #8 | A | Industrial QC, tongue diagnosis | Closed commercial + industry solutions | One fewer grade-A option in vertical vision |
| Step-3 | #17 | B | The fullest four-modality mid-price interface | Closed commercial API | A gap in mid-price modality combinations |
| NeuroXess-N1 | #30 | C | Neural-signal decoding | Project licensing | The only option vanishes; the scenario goes to zero |
| InternVL 4.0 | #16 | A | Open-source perception 96 | MIT weights + official API | The open camp’s perception ceiling drops 6 points |
| Keye-VL-2.0 | #14 | A | Video 94 at ¥2 | Open source (community licence) | The video value anchor disappears |
| LongCat-Next | #9 | A | Generation 92 + research licence | Open source (research licence) | The open-source generation reference disappears |
| WITA-Omni | #23 | A | N4 native + embodied | Closed, filed | The native spectrum loses its top grade |
| Huawei Pangu 5.5 | #21 | B | Industry delivery (typhoon, mining) | Industry licensing | Two industry scenarios revert to bespoke projects |
| MiniCPM-V 4.5 | #26 | A | Edge at ¥0.45 | Apache-2.0 (registration) | The edge price anchor disappears; bills double |
| Hikvision Guanlan 3 | #27 | C | Surveillance video (IR / perimeter / cross-camera) | Industry licensing | Security-video understanding goes to zero; back to traditional CV |
| Cloudwalk Congrong 2 | #32 | B | Recognition (faces, IDs) at perception 91 | Closed commercial | One fewer steady mid-price recognition option |
| Zidong Taichu 2.0 | #25 | B | Academic all-modality ablations | Research licence | The public reference for all-modality alignment disappears |
| Ant Ling 2 | #22 | B | Financial instruments / Chinese 92 | Closed commercial | Finance scenarios lose a tier of Chinese options |
| JD Yanxi 2 | #28 | B | Product-image / SKU corpora | Closed commercial | E-commerce scenarios revert to manual annotation |
| NetEase Fuxi 2 | #29 | B | Game / content generation 82 | Closed commercial | Content-oriented generation loses a low-price option |
Fifteen rows, fifteen different “no second places”. That is the entire survival logic of cold models: they do not survive on strength, they survive on irreplaceability.
How to use this matrix
Three operational notes, for three reader types:
For buyers: always leave one slot on the shortlist for cold models. The method is to walk your scenario against the “exclusive scenario” column — if your business touches any row there, that model enters the POC regardless of its overall rank. A comparison’s popular slots are a marketing function; the matrix’s exclusive-scenario column is a capability function, and the two coincide only on the general battlefield.
For ranking authors: sampling bias is systematically underestimated. The “Chinese multimodal panorama” from ten flagship names is essentially a map of marketing budgets. This site’s approach extends coverage to four classes — general flagships, specialists, cold models and open source — at the cost of maintenance (nine cold-model dossiers update far less often than flagships) in exchange for coverage. For the 1% of readers who need the cold scenarios, these pages are the only answer on the web, and the only answer fears no comparison.
For overlooked vendors: absence from comparisons is not necessarily a loss. SenseTime’s tongue first, NeuroXess’s neural-signal only, MiniCPM’s edge price anchor — these slots are closed to competitors because the comparison sampling frame cannot reach them. The real risk is the day the exclusive scenario is covered by general capability (Step-3 stands exactly on that line), so a cold model’s moat is not the score but the data and delivery records accumulated in the scenario.
Boundaries and method
The matrix’s “exclusive scenario” column is our human judgement from each model’s dossier; the criterion is “does another listed model with comparable capability exist for this scenario”, not market research. Evidence criteria are on the methodology page; per-model scores, benchmarks and sources for the fifteen cold models are on their dossier pages. The matrix updates with versions — if a cold model’s exclusive scenario gets covered, we move it out of the matrix and log the reason in the changelog.
FAQ about this essay
Can you trust the scores for cold models?
Check the evidence grade. Of our 32 models, fourteen are grade A, sixteen grade B and two grade C — cold models span all three: InternVL is A (the most community-reproduced), Step-3 is B, and NeuroXess-N1 is C (every number vendor-reported). The rubric and evidence criteria are public on the methodology page, and every grade-C figure carries its source.
Why do other rankings skip these models?
Three structural reasons: no public API or non-standard pricing (NeuroXess, Pangu), low participation in public benchmarks (SenseTime, StepFun), and the most practical one — maintaining a cold model’s dossier costs more than the traffic it brings. Rankings are content products too, and they run on ROI. This site treats coverage itself as the product value, so the arithmetic differs.
