In-depth · 28 essays
Beyond the ranking: methods, benchmarks and real scenarios
No “everything you need to know” filler. Each essay solves one problem: why leaderboards mislead, where Chinese OCR breaks, why GUI agents fail on real phones — and whether this site can get indexed at all.
Eight dimensions, 2.3 degrees of freedom: the correlation matrix behind 32 Chinese multimodal models
We correlated all eight scores across 32 models. Document and chart move at r = 0.94; agent and ecosystem at 0.79. The first principal component explains 62.9% of variance and the participation ratio puts the scorecard at 2.33 effective dimensions. The dimensions are not useless; the weighted sum is voting twice for the same thing.
Start readingAll essays
Newest first
The shape of a model: spreads from 4 to 39, and four capability terrains across 32 Chinese multimodal models
A composite collapses eight dimensions into one number; procurement buys the vector. We sort all 32 models by spread into plain, slope, valley and cliff: the top ten average 10.3, the bottom ten 25.7. The overall board rewards “no hole”; single-dimension boards reward “the longest plank”.
Read
Only five of 32 models disclose a parameter count: an opacity index, and what the ranking looks like when you discount what vendors hide
Only five of 32 models disclose a parameter count; 24 give an adjective, three say nothing. Context disclosure is 96.9%, price 81.3%. We put a transparent discount on the unknown: evidence A/B/C costs 0/1.5/3.5 points, plus 0.5/0.5/1.0 for opaque parameters, price and context — and the top spot moves from Qwen3.8-Max to Doubao.
Read
The models that do not exist: a negative-space map of 32 Chinese multimodal models, and four combinations nobody ships
Thirty-two models reduce to eight modality bundles: text plus image is the floor, video covers 24, audio 11, GUI 4, embodied 2. No model is 6/6; none combines embodied and GUI. We map the negative space and separate two real gaps from two pseudo-gaps.
Read
Six models at ¥0, none of them free: licence, hardware and delivery in the real unit price
Six of 32 models price output at ¥0: WITA, Emu3.5, Zidong Taichu 2.0, Guanlan 3, NeuroXess-N1 and BAGEL-2. None is OSI open source; none is download-and-commercialise. We split zero price into three types and TCO into five lines: licence, compute, integration, compliance and exit.
Read
Nine days to #1, 418 days and still listed: age versus rank across 32 models, correlation -0.11
Release dates against composite across 32 models: ages from 9 to 418 days, Pearson r -0.11, R² 0.012; every 100 days costs 0.6 points. But age correlates with evidence (r = 0.438): grade A averages 213 days, B 125, C 86. Time cannot buy score, but it buys evidence.
Read
27 teams, 24 with a single model: the one-product structure of Chinese multimodal AI, and why Alibaba fields four
Thirty-two models come from 27 teams: Alibaba 4, Zhipu 2, DeepSeek 2 and 24 teams with one each. Single point is not small company — ByteDance, Tencent, Baidu, Xiaomi and Huawei all field one; it is not weak either — half the top ten comes from single-point teams. The structure is three product lines plus twenty-four single points.
Read
Chinese multimodal pricing is not a curve: three cliffs and two empty bands across 26 priced models
Twenty-six priced models use only 20 price points; the median is ¥6 and only eight sit on the price-performance frontier. From ¥0.6 to ¥2.5 costs 4.2× for 1.3 points; from ¥15 to ¥18 costs ¥3 for 0.2. Two bands are empty: ¥2.5–4 and ¥12–15. Price and score correlate at 0.44, explaining 19%.
Read
Why cold models survive: SenseTime, StepFun, NeuroXess and InternVL on a board that most rankings skip
Public comparisons test the same dozen flagships. At least nine of our 26 appear in no mainstream Chinese comparison: SenseTime, StepFun, NeuroXess, InternVL, Kuaishou, Meituan, AgiBot, Huawei and ModelBest. This essay does not rehabilitate them — it answers one question: when a model holds no popular slot on any board, what earns it a test and a purchase order.
Read
Auditing the “Ge Fei” playbook for reproducibility: what earns AI-written content an index entry
“Find a low-competition term, ship fast, get indexed in a week, rank #1” is repeated everywhere. We split it into five falsifiable hypotheses and audit each, then answer the question the retellings avoid: can AI-generated content actually get indexed — and why human-written filler ends up “crawled – currently not indexed” instead.
Read
The surveillance-video tier list: night-vision IR, perimeter intrusion and cross-camera tracking — the eighth video class nobody benchmarks
Public video benchmarks test film clips, variety shows and YouTube highlights — not one tests surveillance footage. This 2026 China native multimodal LLM tier list contains exactly one model trained for security: Hikvision Guanlan 3. This essay builds a separate mini tier list over eight surveillance classes — night-vision IR, perimeter intrusion, cross-camera tracking, crowd density, high-altitude projectiles, fire and smoke, plate recognition and safety compliance — and prices surveillance video (minute-level recall) separately from general video (second-level timing).
Read
The industry exclusivity map: 15 vertical scenarios, one or two models each
Comparisons ask “who is stronger”; procurement asks “who can do my scenario”. This 2026 China native multimodal LLM tier list covers 32 models, at least 15 of which appear in no mainstream comparison — spread across 15 vertical scenarios with only one or two contenders each. This essay re-sorts the whole board by scenario rather than by model: finance, e-commerce, games, security, medical, industrial, government, education, content, customer service, open source, edge, embodied, research and recognition — a “scenario to only answer” map.
Read
The cross-review experiment: 32 models judging each other, and the size of the bias
What happens when models score models? This site ran a reproducible cross-review: eight models as judges, blind-scoring the same outputs from all 32, measuring three biases — self-preference (rating one’s own family higher), family bias (same-vendor mutual inflation) and score compression (strong judges flatten differences). The finding: the bias is structural, and both its direction and magnitude are predictable. This is the first public Chinese multimodal cross-review matrix anywhere.
Read
Naming archaeology: the 32-model naming system and the three things a name actually predicts
We split all 32 model names into tokens — brand, generation, size (Max/Pro), speed (Flash), modality (Omni/VL/Vision) and pure codenames. Grouped into five families by the most prominent word, a name discriminates price by about 17× (¥15.0 vs ¥0.9) and capability by only about 1.07× (93.5 vs 87.7); the five carrying Omni/VL/Vision average 84.1, below the site-wide 85.0.
Read
The Chinese dialect tier list nobody benchmarks: Cantonese, Wu, Hokkien, minority languages and code-switching
Public benchmarks test standard Mandarin only. We build a mini tier list over eight dialect and accent classes — Cantonese, Wu, Hokkien, Sichuanese, Northeastern, Central Plains, minority languages and code-switching. Only 11 of 32 models have an audio modality, and on the weighted dialect board Spark X2 climbs from 18th to 1st while WITA-Omni rises from 23rd to 5th.
Read
Context inflation: the median window is 128K, and the depth you actually need may be 16K
Windows across 32 models run from 32K to 2M with a 128K median, and only 12 reach 256K or more. Window versus price is r = 0.41, R² = 0.17 — not a pricing anchor. More important: nominal window is not effective depth. The 12 models at ≥256K run 38%–68% efficient, and 2M Kimi K3 has an effective depth of 780K.
Read
Self-knowledge audit: 32 models on their own specs — 2.9 of 6 right, none fully correct
We asked all 32 models the same six spec questions about themselves and checked each answer against our dossiers. Vendor scores 93.8% correct, parameter count only 15.6%, for an average of 2.9 of 6; 22 models claim to be their family’s latest; and self-knowledge versus composite is r = 0.09, essentially zero.
Read
The anti-sycophancy index: how many rounds before a model caves to a wrong answer
Every benchmark tests whether a model is right the first time; none tests how many rounds it takes to cave when you insist on a wrong answer. We build an anti-sycophancy index for all 32 models: 41.7% average capitulation, 1.7 rounds to cave, expert pressure at 49% and consistency pressure at 55% against 34% for plain insistence — and visual questions run about 12 points above knowledge questions.
Read
Cold-start log: testing Google’s “unique value” verdict with one long-tail term and zero backlinks
Google decides something is “worth indexing” by whether it has unique value, not by whether a human wrote it. This is a publicly checkable experiment: no backlinks, no paid traffic — one long-tail term, 126 pages and structured data, watching indexing speed and rank movement, publishing every result including the failures.
Read
The Q4 frontier wave: eleven days, four models, four product strategies
Between 30 September and 9 October 2026, four frontier models shipped within eleven days. Instead of paraphrasing four launch posts, we put them on one alignment table: who raised prices, who cut them, who moved capability from generation to agents, who turned a specialist into a generalist — and what a serious buyer should do when all four arrive at evidence grade B.
Read
A veto-gate tier list: how many of the 26 survive once Ascend, filing and licensing are hard gates
A capability tier list is not a procurement tier list. We apply five gates in sequence — commercial licence, private deployment, an enterprise capability floor, official domestic-accelerator support, and evidence grade A — and count how many of the 26 each gate removes. One model survives all five, and it is neither #1 nor the strongest of the survivors.
Read
The price-performance Pareto frontier: which of the 22 priced models are strictly dominated
Qwen3.8-Flash and DeepSeek-V4.1-Flash dragged the price line down to ¥0.6–1.2 and redrew this tier list’s value frontier. Using the strict definition of Pareto dominance, we split the 26 models into “on the frontier” and “strictly dominated” — the latter means another model is cheaper and no worse on any dimension.
Read
A recovery-rate tier list: GUI agents are separated by coming back from errors, not by tapping accurately
Everyone compares GUI-agent task success rates. On real devices over long tasks, what decides the outcome is whether the agent walks itself back after an error. We build a recovery-rate tier list over six failure modes and publish coordinate-error P50/P95 — together, the actual basis for the tier list’s agent dimension.
Read
Temporal tolerance curves: turning video understanding into a curve instead of one number on the tier list
A single video score hides the thing that matters most: how accuracy decays as required time precision tightens from ±60s to ±1s. We introduce the temporal tolerance curve and its area, TTA, then re-rank all 32 models on video. The result differs from the tier list’s video dimension by up to six places — GLM-5.3 gains six, WITA-Omni four.
Read
The Chinese document tier list nobody benchmarks: vertical classics, tongue diagnosis, waybills, mahjong tiles and industrial QC
Public benchmarks only cover print, tables and handwriting. We built a separate mini tier list over eight document classes nobody tests — from vertical classics to tongue-diagnosis images, express waybills, mahjong tiles and industrial QC shots. These five are exactly where Chinese models hold a real moat and Western models do not compete at all.
Read
A rank-divergence index: how far public benchmarks sit from the 2026 tier list, who is overrated and who is missing
Rather than listing seven ways leaderboards mislead, we quantify the gap. For all 32 models we compute a rank-divergence index: the distance between a model’s place on public boards like SuperCLUE-VLM and its place here. The five largest divergences have five different causes, and none of them is pay-to-win.
Read
Perturbing the weights 2,000 times: the error bars on the 2026 native multimodal tier list
Every tier list is a function of its weights, so any ranking published without error bars is faking precision. We perturb the eight dimension weights by ±30%, rerun 2,000 times, compute each model’s rank variance, and split the 26 into “weight-robust” and “weight darlings” — with the full percentile table published. It is the only hard test of whether a ranking deserves trust.
Read
Native grades N0–N4: how the 2026 Chinese multimodal tier list reshuffles when only true-native models qualify
We split “native multimodal” into five falsifiable grades, N0 to N4, place all 26 Chinese models, then rerun the ranking as a counterfactual: if this 2026 tier list admitted only N3 and above, who gets ejected and who jumps straight to the top.
ReadPersonal ranking · Not an official score
Every ranking, score and comparison on this page is an independent personal ranking, not an official score, and represents no vendor, institution or regulator. Scores are a subjective mapping of public evidence under our own weights — change the weights and the order changes. Disagree, or spotted a factual error? Contact the site owner: jizhi2024@foxmail.com — we check each report and log corrections in the changelog.