The shape of a model: spreads from 4 to 39, and four capability terrains across 32 Chinese multimodal models

A rank is a number; a model is a vector

A tier list compresses eight dimensions into one composite, so we talk about models by rank. But procurement does not buy a rank; it buys a capability vector: where the model is high, where it is low, and how low.

This essay does one thing: turn the eight scores of all 32 models into shapes, classify them by spread (max − min), and answer one question — why the top of the overall board is almost all “flat”, while single-dimension champions are almost all “spiky”.

Spread has a useful property: it is completely weight-invariant. Raise the document weight from 18% to 50% and the ranking moves; the spread does not move a point. Shape is closer to an intrinsic property than rank. Rank is a function of weight assumptions; shape is the geometry of the score vector.

Four terrains

Four buckets by spread — descriptive bins, not capability grades:

• Plain: spread ≤ 8. Eight dimensions pressed together; no obvious peak, no obvious hole.
• Slope: spread 9–19. One or two higher dimensions, but still usable as a general model.
• Valley: spread 20–29. A clear peak and a clear trough; usually needs a partner model.
• Cliff: spread ≥ 30. One dimension near the top, another near unusable; it is a component, not a whole machine.

#	Model	Overall	Spread	Highest	Lowest	Terrain
1	Qwen3.8-Max	95.4	7	Perception 97	Generation 90	Plain
2	Doubao Seed 2.0 Pro	95.2	10	Ecosystem 98	Generation 88	Slope
3	Qwen3.5-397B	94.6	8	Ecosystem 98	Generation 90	Plain
4	GLM-5.3	92.4	12	Ecosystem 97	Generation 85	Slope
5	Kimi K3	92.3	11	Perception 95	Generation 84	Slope
6	GLM-5	92.0	7	Document 95	Video 88	Plain
7	MiMo-V2.5	90.1	7	Video 93	Generation 86	Plain
8	SenseNova 6.5 Pro	89.9	12	Perception 94	Ecosystem 82	Slope
9	LongCat-Next	89.2	22	Document 94	Ecosystem 72	Valley
10	Qwen3-Omni-30B	88.9	7	Document 91	Agent 84	Plain
11	ERNIE 5.1	88.1	11	Chinese 93	Agent 82	Slope
12	DeepSeek-V4.1-Flash	87.9	23	Document 95	Generation 72	Valley
13	Qwen3.8-Flash	87.5	17	Ecosystem 97	Generation 80	Slope
14	Keye-VL-2.0	86.0	16	Video 94	Generation 78	Slope
15	Hunyuan-Vision	85.2	9	Document 89	Agent 80	Slope
16	InternVL 4.0	85.0	30	Perception 96	Video 66	Cliff
17	Step-3	84.7	10	Perception 88	Ecosystem 78	Slope
18	Spark X2	84.5	11	Chinese 90	Agent 79	Slope
19	MiniMax M2.5	83.8	4	Agent 86	Video 82	Plain
20	DeepSeek-OCR 2	81.9	39	Document 99	Generation 60	Cliff
21	Pangu 5.5	81.6	17	Document 87	Generation 70	Slope
22	Ant Ling 2	80.8	32	Chinese 92	Video 60	Cliff
23	WITA-Omni	80.6	32	Video 92	Ecosystem 60	Cliff
24	Emu3.5	80.5	28	Generation 98	Agent 70	Valley
25	Zidong Taichu 2.0	80.1	23	Perception 88	Ecosystem 65	Valley
26	MiniCPM-V 4.5	79.6	14	Document 86	Agent 72	Slope
27	Guanlan 3	78.2	32	Perception 92	Generation 60	Cliff
28	JD Yanxi 2	78.0	34	Chinese 89	Video 55	Cliff
29	Fuxi 2	77.5	20	Chinese 88	Chart 68	Valley
30	NeuroXess-N1	77.2	26	Chinese 88	Ecosystem 62	Valley
31	BAGEL-2	76.9	22	Generation 90	Agent 68	Valley
32	Congrong 2	76.1	26	Perception 91	Generation 65	Cliff

Distribution: 6 plain, 12 slope, 8 valley, 6 cliff. MiniMax M2.5 is the flattest at 4; DeepSeek-OCR 2 is the most jagged at 39 — nearly a 10× difference.

Counterintuition one: the overall board rewards “flat”

The top ten average a spread of 10.3; the bottom ten (ranks 23–32) average 25.7. The higher the rank, the flatter the shape.

This is not a coincidence; it is the arithmetic of a weighted sum. With positive weights on all eight dimensions, one extreme low score drags the total down, while one extreme high score only recovers its own weight. A 99/60 model usually loses to a 90/90 model on the composite.

So the overall champion is not the sharpest model; it is the one with no hole. Qwen3.8-Max spreads 7, Doubao 10, Qwen3.5 8 — the top three are plain or near-plain. Their strength is not a 99 somewhere; it is nothing below 80.

Conversely, single-dimension champions sit near the bottom:

• Document #1 DeepSeek-OCR 2 (99) ranks #20 overall, spread 39;
• Generation #1 Emu3.5 (98) ranks #24, spread 28;
• Video #1 Doubao (96) ranks #2, spread 10 — the closest to plain among single-dimension champions;
• Perception #1 is shared by Qwen3.8-Max and Doubao, spreads 7 and 10, both near-plain;
• Chinese #1 and ecosystem #1 are Doubao (slope); ecosystem co-#1 Qwen3.5 is plain.

The overall board and single-dimension boards are different contests. The overall board rewards “no hole”; single-dimension boards reward “longest plank”. A review that only publishes the overall board systematically hides the value of cliff models; one that only publishes single-dimension boards systematically recommends components that cannot work alone. Read both to know whether you are buying a machine or a part.

Cliff models: six components

Six models have a spread ≥ 30, each with a near-perfect peak and a near-unusable trough:

DeepSeek-OCR 2 (39): document 99 / generation 60. The sharpest component on the board. Its document lead is not “slightly better” — 2 points over #2, 4 over #5 — but generation 60 means no creative output. You are buying a scanner, not an assistant.

JD Yanxi 2 (34): Chinese 89 / video 55. The retail Chinese and product-image ability is real, but video 55 is the lowest here. Use it for live-stream clipping and it fails immediately.

Ant Ling 2 (32): Chinese 92 / video 60. Finance and compliance corpora buy Chinese strength at the cost of near-zero video. Its job is text plus receipts, not general multimodality.

WITA-Omni (32): video 92 / ecosystem 60. Embodied-native video timing paired with the lowest ecosystem score. No public API; the procurement entry is project-based. Comparing it as a token API is a category error.

Guanlan 3 (32): perception 92 / generation 60. Surveillance-specific vision; generation is irrelevant. It sells industry authorisation, not tokens.

InternVL 4.0 (30): perception 96 / video 66. The open-source perception ceiling, with video as a priced-in weakness. It is the best “eye” in an open stack, not a complete multimodal brain.

What these six share: all rank below #16 overall, yet all reach the top ten on their long dimension (most in the top seven). They are not bad models; they are single-function models. Buying them off the overall board is a mistake; buying them off the single-dimension board is correct.

Counterintuition two: three inverted peaks

Most models are “understanding high, generation low”: perception, documents and agents score above generation. Four flip it:

Model	Generation	Understanding average (perception/document/chart/video)	Generation − understanding
Emu3.5	98	81.5	+16.5
BAGEL-2	90	77.5	+12.5
Fuxi 2	82	77.0	+5.0
LongCat-Next	92	91.0	+1.0 (near-inverted)

The first three are genuinely “can draw, cannot read”. Emu3.5 has the board’s best generation at 98, but perception 86, document 82, video 78 and agent 70; BAGEL-2 has generation 90 and agent 68. They are creative tools, not analysis tools.

An inverted peak is not a defect; it is positioning. If you need “turn a brief into a visual set”, Emu3.5 fits better than most of the top five. If you need “read an invoice and fill a system”, its document 82 disappoints. Shape tells you which end it serves; the total does not.

Same score, different shape: three pairs

Close totals with opposite shapes are where the overall board misleads most:

LongCat-Next (89.2, spread 22) vs Qwen3-Omni-30B (88.9, spread 7). A 0.3-point difference, but LongCat is a double peak — document 94, generation 92, ecosystem 72 — while Omni is a plain of 84–91. For a general assistant, take Omni; for document-plus-generation with an ecosystem hole, take LongCat.

Pangu 5.5 (81.6, spread 17) vs MiniMax M2.5 (83.8, spread 4). MiniMax sits at 82–86 everywhere; Pangu is document 87, generation 70. MiniMax wins the total, but if you only need documents, Pangu’s 87 is better; if you need a model that does not flinch at any task, MiniMax’s “no hole” is worth more.

InternVL 4.0 (85.0, spread 30) vs Hunyuan-Vision (85.2, spread 9). Nearly identical totals: one is an open-source eye (perception 96 / video 66), the other a closed generalist (80–89 everywhere). The first fits a self-built perception stack; the second fits direct API calls.

A total is a scalar; procurement is a vector. Two models with the same total can be usable and unusable in the same scenario.

Shape is more stable than rank

The weight-perturbation experiment shows that perturbing the eight weights by ±30% and rerunning 2,000 times moves the middle and tail a lot; some models drift across five or six places.

Spread does not move. Weights change the relative importance of dimensions, not the dimension scores. Raise video from 14% to 40% and DeepSeek-OCR 2 still spreads 39; cut it to 4% and it still spreads 39. Rank is a function of weight assumptions; shape is a property of the score vector.

That suggests a practical selection order:

1. Read shape first: are you buying a machine or a part? Plain/slope for machines, valley/cliff for parts.
2. Then read the long dimension: for a part, is the peak exactly your scenario?
3. Then read rank: among shape-compatible candidates, sort by the total or your own weights.

Reverse the order — rank first, shape second — and the board’s bias toward balance filters out every component before you see it, leaving you wondering why no document-99 model exists.

Procurement reading

Shape	Buy it for	Do not buy it for	Examples
Plain	General assistant, single-vendor standard, support/knowledge base	A pipeline needing one extreme	Qwen3.8-Max, Qwen3.5, MiniMax
Slope	General work with one stronger area	A 99-point single step	Doubao, GLM-5.3, Kimi
Valley	Two or three specialities, paired with another model	Running a whole workflow alone	LongCat, Emu3.5, DeepSeek-V4.1
Cliff	One step in a pipeline	A whole machine or general assistant	OCR 2, InternVL, Guanlan, Yanxi

A simple rule: if you cannot say in one sentence which step of your system this model owns, do not buy a cliff model. All of its value comes from that step; outside it, the total is meaningless.

Boundaries

• Spread is the geometry of eight scores, not an architectural property. Two models with the same spread may have completely different architectures.
• Spread ignores which dimension is high. LongCat and BAGEL both spread 22, but one peaks on document plus generation, the other on generation; read the highest/lowest columns together.
• Scores carry error. Grade B is ±1.5 points, grade C wider; under error a spread can move 2–7 points. The ordering of broad terrains is robust; adjacent ranks are not.
• Eight dimensions are coarse. The real shape may be more complex; this is the finest version public data supports.

Reproduction

Spread = highest of the eight scores minus the lowest, computed directly from src/data/models.ts, with no weights and no prices. The table is sorted by the composite, recomputed with the site weights. Scripts ship with the data; rerun after every expansion, and the shape table updates in the changelog.

If this essay leaves one sentence: the overall board buys “no hole”; single-dimension boards buy “the longest plank”. Know whether you are buying a machine or a part before you read the rank.