27 teams, 24 with a single model: the one-product structure of Chinese multimodal AI, and why Alibaba fields four

32 models, 27 teams

Deduplicate the 32 listed models by vendor and you get 27 teams. The distribution is very uneven:

• Alibaba, 4 models: Qwen3.5-397B (#3), Qwen3-Omni-30B (#10), Qwen3.8-Max (#1), Qwen3.8-Flash (#13);
• Zhipu AI, 2: GLM-5 (#6), GLM-5.3 (#4);
• DeepSeek, 2: DeepSeek-OCR 2 (#20), DeepSeek-V4.1-Flash (#12);
• the other 24 teams, one each.

In other words, 24 of 27 teams stake their entire multimodal bet on a single model. Those 24 “single points” include giants — ByteDance, Tencent, Baidu, Xiaomi, Huawei — as well as model companies such as Moonshot, SenseTime, Meituan, Kuaishou, iFlytek, StepFun and MiniMax, and industry/research teams such as Hikvision, Cloudwalk, NeuroXess, AgiBot, ModelBest, Shanghai AI Lab, CASIA, Ant, JD and NetEase.

Single point is not a small-company trait; it is a product strategy. This essay rereads the board through “product-line density”: who uses one model to hit one point, who uses several models to cover a surface, and what that means for procurement.

The three multi-product teams

Alibaba: four models, four price bands

Model	Overall	Score	Price / 1M output tokens	Modalities	Role
Qwen3.8-Max	#1	95.4	¥18	T/I/V/A/GUI	Flagship, capability ceiling
Qwen3.5-397B	#3	94.6	¥12	T/I/V/A	Previous flagship, value
Qwen3-Omni-30B	#10	88.9	¥4	T/I/V/A	Open all-modality (Apache-2.0)
Qwen3.8-Flash	#13	87.5	¥1.2	T/I/V/GUI	Low-cost GUI/agent entry

Four models cover ¥1.2, ¥4, ¥12 and ¥18, and five of six modalities (everything except embodied). Alibaba is not competing at one point; it is surrounding four.

The internal staircase is worth reading:

• Max → Qwen3.5: −0.8 points, −33% price. The flagship premium is thin; Qwen3.5 is the “sweet spot” sub-flagship.
• Qwen3.5 → Omni: −5.7 points, −67% price. Omni does not sell general capability; it sells “all-modality at ¥4 plus Apache-2.0”.
• Omni → Flash: −1.4 points, −70% price. Flash is 70% cheaper than Omni for 1.4 fewer points, but loses audio; it sells the “¥1.2 GUI/agent ticket”.

Four models are not four repeats; they are four ecological niches. That is the first advantage of a multi-product team: different price bands without making one model compromise across all of them.

Zhipu: 5.3 dominates 5

GLM-5.3 (#4, 92.4, ¥9) versus GLM-5 (#6, 92.0, ¥10):

• the newer version is +0.4 points;
• it is ¥1 cheaper;
• both are closed commercial plus private deployment, with 5.3 adding “Ascend full-stack training/images”.

From the public dossier, GLM-5 is dominated by GLM-5.3 on both score and price. Its remaining reason to exist can only be found in existing contracts, certifications and delivery cycles, not in the capability table.

That is the second trait of multi-product teams: self-cannibalisation. The newer version is stronger and cheaper, so the old one becomes harder to sell; but it cannot be withdrawn immediately because customers, tender specifications and private deployments already exist. For buyers: when a sibling iteration appears, first ask whether the old version still has a reason to exist, then decide whether to pay a premium for “stability”.

DeepSeek: not a product line, two nails

DeepSeek-OCR 2 (#20, 81.9, ¥0.8, document 99) and DeepSeek-V4.1-Flash (#12, 87.9, ¥0.6, document 95):

• one specialises in document parsing, spread 39;
• one is a cheap generalist, spread 23;
• both are under ¥1 and neither is a flagship.

DeepSeek’s two are not a product line; they are two steps in a pipeline. OCR 2 does first-pass layout parsing; V4.1-Flash does second-pass review and structuring; the combined cost per thousand images is far below “flagship review” (see cost per 1,000 images).

A third form of multi-product team: not price-band coverage, not generational iteration, but task division.

The 24 single-point teams

Team	Model	Overall	Price	Type
ByteDance	Doubao Seed 2.0 Pro	#2	¥15	General
Moonshot AI	Kimi K3	#5	¥16	General/long context
Xiaomi	MiMo-V2.5	#7	¥6	General
SenseTime	SenseNova 6.5 Pro	#8	¥12	Vision/industry
Meituan	LongCat-Next	#9	¥2.5	Native multimodal
Baidu	ERNIE 5.1	#11	¥8	General/search
Kuaishou	Keye-VL-2.0	#14	¥2	Video
Tencent	Hunyuan-Vision	#15	¥8.6	General
Shanghai AI Lab	InternVL 4.0	#16	¥1.0	Open perception
StepFun	Step-3	#17	¥5	General
iFlytek	Spark X2	#18	¥6	Speech/education
MiniMax	MiniMax M2.5	#19	¥8	Agent/general
Huawei	Pangu 5.5	#21	¥20	Industry
Ant Group	Ant Ling 2	#22	¥5	Finance
AgiBot	WITA-Omni	#23	¥0	Embodied
BAAI	Emu3.5	#24	¥0	Generation/world model
CASIA	Zidong Taichu 2.0	#25	¥0	Academic all-modality
ModelBest	MiniCPM-V 4.5	#26	¥0.5	Edge
Hikvision	Guanlan 3	#27	¥0	Surveillance
JD Group	JD Yanxi 2	#28	¥6	E-commerce
NetEase	Fuxi 2	#29	¥4	Games/content
NeuroXess	NeuroXess-N1	#30	¥0	Neural signals
ByteDance · PKU	BAGEL-2	#31	¥0	Open unified model
Cloudwalk	Congrong 2	#32	¥7	Recognition

Three counterintuitive facts about these 24:

Single point does not mean small company. ByteDance, Tencent, Baidu, Xiaomi and Huawei each field one model; their multimodal investment is not small, they simply chose one model for one main scenario. Product-line density is strategy, not scale.

Single point does not mean weak. Doubao #2, Kimi #5, MiMo #7, SenseNova #8, LongCat #9 — half of the top ten comes from single-point teams. One model done extremely well beats four models stuck in the middle.

Single-point models are spikier. The 24 single-point models average a spread of 19.1; the eight multi-product models average 15.0. Eight of 24 single-point models (33%) spread ≥25, versus one of eight multi-product models (DeepSeek-OCR 2, 12.5%). Single-point teams more often use one model to cover one scenario, so the shape is sharper; multi-product teams can divide labour across models, so the portfolio is flatter.

Why “single point + three portfolios”

This is not accidental; it follows from three constraints:

1. Multimodal training is expensive. A top-20 model needs a vision encoder, a language base, and joint training on video, audio and agent data; maintaining four product lines at once requires compute and data only a few teams have.

2. Scenario data is vertical. Surveillance, finance, e-commerce, games, medical and neural signals each collect data differently. A team with first-hand data in one scenario beats a team with second-rate data in four. Single point is a data strategy, not a capability deficit.

3. Commercial paths differ. General models sell APIs, industry models sell authorisation/projects, open models sell ecosystem. The organisational capabilities are completely different; doing all three costs more management than it returns.

So the market splits into two layers:

• three multi-product teams (Alibaba, Zhipu, DeepSeek) covering price, task and generation with product lines;
• twenty-four single-point teams using one model to hit one point, surviving on scenario depth or cost.

This is not consolidation; it is a long tail. 27 teams, 32 models, 1.19 models per team on average; the median is one. The market has not converged into a few full-stack giants; it has grown into a field of single points.

Procurement: different risks for single point and multi-product

Dimension	Single-point team	Multi-product team
Upgrade path	Next generation may rename/re-architect; migration uncertain	Sibling staircase, smoother upgrade
Single point of failure	One model withdrawn = supply interruption	One withdrawn, swap within family
Self-cannibalisation	None (only one model)	Yes (new and old replace each other)
Roadmap	Focused, fast iteration	Diffuse, resources may tilt
Negotiation	One model, little leverage	Bundle, more leverage
Integration	One model, simple	Multiple models, orchestration complexity

The single-point risk is “no spare tyre”; the multi-product risk is “too many spare tyres, which one?”

Ask two questions in procurement:

1. For a single-point team: if this model is replaced in 12 months, what is the migration path? Can weights, API and data leave?
2. For a multi-product team: is there a cheaper, stronger sibling covering the same scenario? If so, why buy this one?

The second question already has an answer at Zhipu: GLM-5.3 dominates GLM-5. For a multi-product vendor, deduplicate within the family first, then compare across families.

A prediction and its falsification conditions

Prediction: in the next 12 months, Chinese multimodal will not converge into “three full-stack vendors”; it is more likely “3–5 multi-product teams + a field of single points”, because scenario data remains vertical and training remains expensive.

Falsification conditions: this prediction is wrong if any of the following happens —

• the number of single-point teams falls below 15 (M&A, exit, absorption by platforms);
• one of the three multi-product teams expands beyond eight models and covers more than five scenarios;
• the average spread of single-point models falls from 19.1 to below 12 (they are filling gaps rather than deepening scenarios).

All three can be recomputed after future expansions. Predictions are not written on a wall; they are written in data.

Boundaries

• Vendor grouping uses the vendor field. “ByteDance · PKU” BAGEL-2 and “ByteDance” Doubao count as two teams; merging them makes 23 single-point teams. Definitions change numbers.
• Product-line density is not capability. Alibaba’s four do not make Alibaba the strongest; one model does not make a team weak. Density describes strategy, not score.
• “Single-point models are spikier” is description, not causation. Single-point teams may also produce sharper models because resources are concentrated; multi-product teams may be flatter because of division of labour. The sample is 27 teams.
• The prediction is judgement. The falsification conditions are written here for a 12-month look-back.

Reproduction

Group by the vendor field; count models, average rank and average spread per group; single-point = one model, multi-product = more than one. Data from src/data/models.ts; scripts ship with the data; rerun after every expansion, and the team table updates in the changelog.

If this essay leaves one sentence: Chinese multimodal is not “a few taking all”; it is “three product lines plus twenty-four single points”. The default procurement shape is not one full-stack vendor, but a map of multiple vendors.