Three things to remember
- The gates cut 32 → 28 → 11 → 6 → 2 → 1. Gate two, private deployment, removes seventeen at once — including the tier list’s top three: Qwen3.8-Max, Doubao Seed 2.0 Pro and Qwen3.5-397B. In 2026 the capability frontier and the self-hostable frontier run in opposite directions.
- The two new models die at two different gates: InternVL 4.0 at the capability floor (agent 78 < 84), and NeuroXess-N1 at weight custody (project licensing — weights never leave for the client).
- DeepSeek-OCR 2 — first on documents site-wide at 99, and ¥0.8 per thousand images — dies at gate three because its agent score is only 72. A weighted average lets documents compensate for agents; an enterprise workflow does not. That is exactly where the composite stops working for procurement.
The top three of the tier list cannot be bought in a real procurement
This is the most uncomfortable essay on this site.
The top three of the 2026 China native multimodal LLM tier list are Qwen3.8-Max (95.4), Doubao Seed 2.0 Pro (95.2) and Qwen3.5-397B (94.6). If you are an enterprise that needs data to stay on-prem, runs on domestic compute, and must go through public tender, all three are eliminated — and the reasons have nothing to do with capability.
The composite is a weighted average; it assumes every dimension can compensate for another. But enterprise procurement is not a weighted average, it is one-vote veto: fail one hard gate and full marks on the other seven are irrelevant. So we take the tier list apart and turn it into serial gates, watching each gate remove whom.
Five gates, none negotiable
| Gate | Criterion | Source | Nature |
|---|---|---|---|
| G1 | Commercially licensable (research licences out) | Licence field | Objective |
| G2 | Privately deployable / weights self-held | Open-source flag + licence terms | Objective |
| G3 | Enterprise capability floor: document ≥ 90 and chart ≥ 85 and agent ≥ 84 | Our scores | Threshold subjective, scores public |
| G4 | Official domestic-accelerator support (public image or adaptation statement) | Vendor public materials | Our judgement, disputable |
| G5 | Evidence grade A (third-party reproducible) | Our evidence grading | Our judgement, disputable |
G1 and G2 are hard facts written in licences. G3’s scores are public and its thresholds are ours — change them if you like. G4 and G5 are our judgements with the standards laid open; you may disagree.
The five gates are ordered, and the order itself produces different conclusions. We run a sensitivity analysis at the end.
G1 commercially licensable: 32 → 28
A research licence is not a commercial licence — the most overlooked gate of all. Four models exit:
| Out | Composite | Overall | Licence |
|---|---|---|---|
| LongCat-Next | 89.2 | #9 | Open source (research licence) |
| Emu3.5 | 80.5 | #24 | Open source (research licence) |
| Zidong Taichu 2.0 | 80.1 | #25 | Open weights (research licence) |
| BAGEL-2 | 76.9 | #31 | Open source (research licence) |
LongCat-Next hurts most here: 9th overall, 89.2, ¥2.5 output price, third point on the Pareto frontier, generation 92 second on the site. Its research licence means you cannot even deploy internally in production — not “licensing to be negotiated”, the terms exclude it outright. Plenty of teams prove it out in POC and only discover at procurement that it cannot be used — the most common rework of 2026 Chinese open-source models.
Zidong Taichu 2.0 is this cycle’s new casualty on the same grounds: the academic model’s weights are open under a research licence — citable in papers, unusable in production. It, Emu3.5 and BAGEL-2 all matter less in practice: all three are generation- or research-oriented and rarely mainstays in enterprise flows.
G2 privately deployable: 28 → 11, the top three wiped out
This gate removes seventeen, including the strongest on the capability board:
| Out | Composite | Overall | Why |
|---|---|---|---|
| Qwen3.8-Max | 95.4 | #1 | Closed commercial API; flagship weights unavailable |
| Doubao Seed 2.0 Pro | 95.2 | #2 | Closed commercial API |
| Qwen3.5-397B | 94.6 | #3 | Partial sizes open, 397B flagship closed |
| MiMo-V2.5 | 90.1 | #7 | Closed API + partial open source |
| SenseNova 6.5 Pro | 89.9 | #8 | Closed commercial |
| ERNIE 5.1 | 88.1 | #11 | Closed commercial |
| Hunyuan-Vision | 85.2 | #15 | Closed commercial |
| Step-3 | 84.7 | #17 | Closed commercial |
| Spark X2 | 84.5 | #18 | Closed commercial |
| MiniMax M2.5 | 83.8 | #19 | Closed commercial |
| WITA-Omni | 80.6 | #23 | Closed, filed for commercial use, preview |
| NeuroXess-N1 | 77.2 | #30 | Closed · project licensing; vendor deploys, weights never leave for the client |
| Hikvision Guanlan 3 | 78.2 | #27 | Closed · industry licensing; ships with hardware, no client weight custody |
| Ant Ling 2 | 80.8 | #22 | Closed commercial; no on-prem licence clause |
| JD Yanxi 2 | 78.0 | #28 | Closed commercial; no on-prem licence clause |
| NetEase Fuxi 2 | 77.5 | #29 | Closed commercial; no on-prem licence clause |
| Cloudwalk Congrong 2 | 76.1 | #32 | Closed commercial; no on-prem licence clause |
The top three of the 2026 China native multimodal LLM tier list all exit at this gate. The three strongest models are precisely the three pure cloud APIs.
That is not coincidence — it is structural: in 2026 the capability frontier and the self-hostable frontier run in opposite directions. Training a 95-point native multimodal flagship needs trillion-scale parameters and the newest compute, and vendors have no incentive to release it; those willing to open weights are either mid-sized or last-generation flagships.
Qwen3.5-397B’s row deserves its own note: its licence reads “partial sizes open (Apache-2.0), flagship closed”. That is a selection trap — you can download the Qwen3.5 family from Hugging Face, but not the one ranked #3 here. Same-generation size gaps reach 7+ points in our scoring.
Two “deploys inside the client network” models die the same way: NeuroXess-N1 is project-licensed — the model runs in the client’s machine room while weights and legal control never do; Guanlan is industry-licensed — shipped with hardware and solutions, again with no client weight custody. Under tender language both fail G2: what you need is not “data stays on-prem” but “assets are self-held” — the first is a compliance posture, the second is the procurement object. This cycle’s four closed commercial entries (Ling 2, Yanxi 2, Fuxi 2, Congrong 2) die for a simpler reason: the terms contain no on-prem licence at all — the procurement entry for industry models is the solution, not the weights.
Eleven survive:
| Rank | Model | Composite | Tier | Evidence | How weights are held |
|---|---|---|---|---|---|
| #4 | GLM-5.3 | 92.4 | S | B | Closed commercial + on-prem licence (Ascend images included) |
| #5 | Kimi K3 | 92.3 | S | A | Open weights (community licence) |
| #6 | GLM-5 | 92.0 | S | A | Closed commercial + on-prem licence |
| #10 | Qwen3-Omni-30B | 88.9 | A | A | Apache-2.0 |
| #12 | DeepSeek-V4.1-Flash | 87.9 | A- | B | Open weights (MIT) |
| #13 | Qwen3.8-Flash | 87.5 | A- | B | Open weights (Apache-2.0) |
| #14 | Keye-VL-2.0 | 86.0 | A- | A | Open source (community licence) |
| #16 | InternVL 4.0 | 85.0 | A | A | Open weights (MIT) |
| #20 | DeepSeek-OCR 2 | 81.9 | B+ | A | Open source (MIT-family) |
| #21 | Huawei Pangu 5.5 | 81.6 | B+ | B | Closed, industry licensing |
| #24 | MiniCPM-V 4.5 | 79.6 | B | A | Apache-2.0 (commercial registration required) |
Note the tier gap in this table: from #6 at 92.0 it jumps straight to #10 at 88.9, 3.1 points of empty space. In other words, once you require private deployment, your capability ceiling is locked at 92.4, 3 points below the 95.4 available without the gate. That 3 points is the real price of privatisation, and it has nothing to do with money.
G3 capability floor: 11 → 6, six new casualties
The main battlefield of enterprise multimodality is three things: reading documents, reading charts, calling tools. We set the floor at document ≥ 90, chart ≥ 85, agent ≥ 84.
| Out | document | chart | agent | Where it fails |
|---|---|---|---|---|
| DeepSeek-OCR 2 | 99 | 88 | 72 | agent |
| Keye-VL-2.0 | 86 | 85 | 83 | document, agent |
| Huawei Pangu 5.5 | 87 | 84 | 74 | All three |
| MiniCPM-V 4.5 | 86 | 80 | 72 | All three |
| InternVL 4.0 | 93 | 90 | 78 | agent |
| Ant Ling 2 | 88 | 82 | 80 | All three, each by one or two points |
| JD Yanxi 2 | 86 | 78 | 78 | All three |
| NetEase Fuxi 2 | 70 | 68 | 75 | All three |
| Cloudwalk Congrong 2 | 78 | 72 | 70 | All three |
DeepSeek-OCR 2 is the most surprising casualty at this gate. Its document score of 99 is first among all 32 models on the tier list, two points above Qwen3.8-Max and Qwen3.5-397B, at ¥0.8 per thousand images. On any “document intelligence” board it is the champion.
But its agent score is 72, 23rd of 26. It is a pure information-extraction engine, not a model that can participate in a business process. Wire it into an OCR pipeline and it is the best choice; make it the QA brain of an enterprise knowledge base and it stalls wherever multi-step tool calls are needed.
That is where weighted averages mislead: in the composite, document 99 (weight 18%) compensates for most of agent 72 (weight 14%), so it lands at 20th and looks “fine”. In an enterprise scenario, the two simply cannot compensate for each other.
This cycle’s four industry models all die at G3, each in a different shape. Ant Ling 2 misses all three by one or two points — the “most regrettable” casualty; Yanxi 2 fails all three; Fuxi 2’s document 70 shows it was never designed for enterprise flows; Congrong 2 fails all three at the lowest scores. What they share: however deep the scenario corpora, they do not buy process capability — enterprise flows test three general things, and scenario specialisation earns no points on any of them.
InternVL 4.0 is this cycle’s new casualty, with the same cause as DeepSeek-OCR 2 but a different shape. Its document 93 and chart 90 clear the bar; agent 78 misses by 6. It is an excellent perception component — perception 96, MIT licence, grade-A evidence — but enterprise flows need participants who can finish three steps, not components. Open-source models have caught up on point capability and are still a generation behind on process capability — that is the most stable shape of the G3 wall.
Huawei Pangu 5.5 fails all three, the direct cause of its 21st composite. Its value is in industry delivery cases, not in the general capability floor.
Six survive: GLM-5.3, Kimi K3, GLM-5, Qwen3-Omni-30B, DeepSeek-V4.1-Flash, Qwen3.8-Flash.
G4 domestic accelerators: 6 → 2, four open-weight models die at “no official image”
Criterion: does the vendor provide an official inference image or public adaptation statement for domestic accelerators (Ascend / Cambrian / Hygon)?
| Out | Why |
|---|---|
| Kimi K3 | Open weights (community licence); no official domestic-chip image |
| Qwen3-Omni-30B | Apache-2.0; no official domestic-chip image |
| DeepSeek-V4.1-Flash | Open weights (MIT); no official domestic-chip image |
| Qwen3.8-Flash | Apache-2.0; no official domestic-chip image |
This is the most disputable gate here, and we spell out the dispute.
All four casualties are open-weight. Technically your team can port them to Ascend — MIT and Apache-2.0 present no legal obstacle. So what this gate blocks is not “can it be done” but “who owns the porting cost and adaptation responsibility”.
Under tender language those are entirely different things: the buyer wants a vendor-issued adaptation statement and after-sales responsibility, not “portable in theory”. A self-ported model running on domestic chips has no vendor to hold accountable when something breaks, and no official document to cite in audit. So from a procurement-compliance angle the gate is real; from a technical-feasibility angle it is virtual.
If your organisation has a strong inference team and compliance accepts self-owned adaptation responsibility, skip G4 and your candidate set is 6 models, not 2 — with Qwen3.8-Flash and DeepSeek-V4.1-Flash’s price advantages (¥1.2 / ¥0.6) cutting overall TCO by an order of magnitude. That is a real, worthwhile trade, not a compromise.
The two that pass G4:
| Model | Composite | Evidence | Adaptation |
|---|---|---|---|
| GLM-5.3 | 92.4 | B | Licence explicitly states “Ascend images included”; full-stack Ascend training |
| GLM-5 | 92.0 | A | Zhipu’s public Ascend adaptation statement; multiple public-sector deliveries |
G5 evidence grade: 2 → 1, the one eliminated is the higher-scoring one
Final gate: evidence grade A — our scoring is backed by third-party independent reproduction.
- GLM-5: grade A. Passes.
- GLM-5.3: grade B. Out. Released 2026-10-02, under three weeks at the time of writing; the SuperCLUE-VLM 2026-10 score of 89.4 is vendor-reported, as is SWE-Bench Pro 73%; no third-party reproduction yet.
The final answer is a model ranked 6th overall at 92.0: GLM-5.
And GLM-5.3 — 0.4 points higher, newer across the board, clearing the first four gates — is eliminated by “certainty”.
That looks absurd, but it is the necessary product of this site’s entire scoring system, and the reason we insist on publishing evidence grade separately: capability risk and certainty risk are two different risks and cannot be expressed in one number.
- GLM-5.3’s risk: the 92.4 might be 91, might be 93.5 — we do not know.
- GLM-5’s risk: the 92.0 we are comparatively sure of.
For an enterprise that must commit to an SLA in contract, the second risk is priceable and the first is not. But if your scenario is internal tooling that can tolerate one model swap, GLM-5.3’s 0.4-point edge and native Ascend images are worth the risk.
That is why G5 goes last, rather than inside the composite.
Sensitivity analysis: does changing one threshold change the conclusion?
It does. That must be said.
| Change | Final candidates | Count |
|---|---|---|
| Baseline (all five gates) | GLM-5 | 1 |
| Drop G5 (no grade-A requirement) | GLM-5.3, GLM-5 | 2 |
| Drop G4 (self-owned chip adaptation) | + Kimi K3, Qwen3-Omni-30B, DeepSeek-V4.1-Flash, Qwen3.8-Flash | 6 |
| G3 agent threshold down to 78 | + InternVL 4.0 (agent 78, just over) | 7 |
| Drop G2 (accept pure cloud API) | + Qwen3.8-Max, Doubao, Qwen3.5-397B and 10 more | 14 |
| Keep G1 only (commercially licensable) | 28 models | 28 |
The most sensitive gate is G2; the least sensitive is G3’s document threshold. That result carries its own information: the real watershed of enterprise selection is deployment form, not capability detail. Haggling over 5 points of document threshold is pointless; deciding once whether to privatise determines the fate of seventeen models.
Note that lowering G3’s agent threshold to 78 admits only one more model — six models crowd between 87 and 78 (including this cycle’s Ling 2 at 80, Yanxi 2 at 78 and Fuxi 2 at 75), and none clears document and chart at the same time. A capability floor is a wall, not a gate: relaxing one dimension rarely changes the pass count, because model capability profiles are skewed.
If you remember one thing: decide G2 first, then read the tier list. Do it the other way and you will spend three weeks evaluating a batch of models you cannot buy at all.
Three ready-to-use shortlists
A. Strong compliance (finance, government, central SOEs: data on-prem + domestic compute + auditable) → GLM-5 (92.0 / S / grade A) — the single option. → Fallback path: wait for GLM-5.3’s third-party reproduction (expected in 1–2 quarters), raising the ceiling from 92.0 to 92.4. → Do not: backfill with Pangu 5.5. It is full-stack domestic but fails all three G3 items.
B. Medium compliance (manufacturing, health IT, large private enterprises: data on-prem, self-owned adaptation accepted) → GLM-5.3 (92.4 / S) as the main, Kimi K3 (92.3 / S / grade A) as the control. → These two are the only combination above 92 with self-held weights. Kimi K3’s grade A is its real edge over GLM-5.3. → For video-heavy scenarios add Qwen3-Omni-30B (video 90, Apache-2.0).
C. Weak compliance (internet companies, internal tools, data egress tolerated) → Read the tier list directly and use Qwen3.8-Max (95.4); the first five sections of this essay do not apply to you. → But note Qwen3.8-Max is grade B with vendor-reported figures not yet reproduced; the same-price Doubao Seed 2.0 Pro (95.2 / grade A) is 0.2 points back with higher certainty. In enterprise procurement, 0.2 points for a grade-A trail is usually a good trade.
Method statement
G4 and G5 are our judgements, not industry standards. G4’s domestic-chip adaptation data comes from vendor public materials, collected 2026-10-18, and may lag the latest adaptation releases. G3’s three thresholds are set by us based on typical enterprise multimodal workloads and carry no authoritative backing — we provide a recomputable framework, not a conclusion. Change thresholds, change gate order, add or remove gates — all of it can be explored alongside this essay’s tables in the compare lab.
Full scoring and evidence criteria are on the methodology page; historical gate adjustments are in the changelog.
FAQ about this essay
Do non-Xinchuang projects need this veto board?
Yes, but only the first three. Commercial licence, private deployment and the capability floor hold for any enterprise project regardless of Xinchuang; gate four, official domestic-accelerator support, is Xinchuang-specific. Gate four also has two readings — strictly by official images you get two candidates; if your inference team will own the port and compliance accepts that, you get six, at an order of magnitude lower TCO.
How often is gate information refreshed?
Filing and licence status track regulation, so we recheck quarterly and log changes in the changelog; chip support and delivery records track vendor releases. When any item moves, the dossier’s evidence grade and gate flags update with it rather than waiting for a major release.
