Full ranking · October 2026

Chinese multimodal tier list: all 32 models ranked

Composite, eight dimension scores, price, context and evidence grades in one table. Sort by any dimension, or filter to open and native models. Gaps under 0.5 count as level — do not agonise over 0.1.

Last updated Oct 18, 2026Editorial scoring

20models

8dimensions

13grade-A evidence

#1S+

Qwen3.8-Max

Alibaba

The flagship that changed who is #1

95.4/ 100

#2S+

豆

Doubao Seed 2.0 Pro

ByteDance

The practical ceiling for Chinese multimodal work

95.2/ 100

#3S+

Qwen3.5-397B-A17B

Alibaba

The multimodal flagship that behaves like infrastructure

94.6/ 100

Tier board

From S+ to B: where the gaps are

The S+/S gap comes from modality coverage and long-task stability. A-tier models often match the leaders in one or two dimensions but carry a visible hole.

S+Category leader3

QMQwen3.8-Max95.4 豆Doubao Seed 2.0 Pro95.2 QQwen3.5-397B-A17B94.6

STier 13

G53GLM-5.392.4 KKimi K392.3 GLMGLM-592.0

A+Strong challenger3

MiMiMo-V2.590.1 SNSenseNova 6.5 Pro89.9 LCLongCat-Next89.2

AReliable workhorse7

QoQwen3-Omni-30B88.9 文ERNIE 5.188.1 DFDeepSeek-V4.1-Flash87.9 QFQwen3.8-Flash87.5 KeKeye-VL-2.086.0 混Hunyuan-Vision85.2 IVInternVL 4.085.0

A-Specialist3

S3Step-384.7 火Spark X284.5 MMMiniMax M2.583.8

B+Viable alternative9

DSDeepSeek-OCR 281.9 PPangu 5.581.6 ALAnt Group Ling 280.8 WWITA-Omni Preview80.6 EEmu3.580.5 ZTZidong Taichu 2.080.1 CPMMiniCPM-V 4.579.6 HGHikvision Guanlan 378.2 JYJD Yanxi 278.0

BWatchlist4

FXNetEase Fuxi 277.5 NXNeuroXess-N177.2 BBAGEL-276.9 CRCloudwalk Congrong 276.1

Data table

The full sortable ranking

Sorted by composite by default. Click Docs, Video or Agent to re-sort by a single dimension — usually closer to a real procurement decision.

#	Model	Tier							Context	Type
1	Qwen3.8-MaxAlibaba	S+	95.4	97	97	93	95	¥18	1M (256K native)	Closed · Native
2	Doubao Seed 2.0 ProByteDance	S+	95.2	97	96	96	93	¥15	128K	Closed · Native
3	Qwen3.5-397B-A17BAlibaba	S+	94.6	96	97	92	92	¥12	256K (up to 1M)	Closed · Native
4	GLM-5.3Zhipu AI	S	92.4	93	94	88	95	¥9	400K	Closed · Native
5	Kimi K3Moonshot AI	S	92.3	95	92	90	95	¥16	2M	Open · Native
6	GLM-5Zhipu AI	S	92.0	93	95	88	91	¥10	200K	Closed · Native
7	MiMo-V2.5Xiaomi	A+	90.1	91	90	93	90	¥6	1M	Closed · Native
8	SenseNova 6.5 ProSenseTime	A+	89.9	94	93	89	86	¥12	128K	Closed · Native
9	LongCat-NextMeituan	A+	89.2	92	94	87	88	¥2.5	256K	Open · Native
10	Qwen3-Omni-30BAlibaba	A	88.9	90	91	90	84	¥4	32K	Open · Native
11	ERNIE 5.1Baidu	A	88.1	90	92	85	82	¥8	128K	Closed · Native
12	DeepSeek-V4.1-FlashDeepSeek	A	87.9	90	95	78	85	¥0.6	128K	Open · Native
13	Qwen3.8-FlashAlibaba	A	87.5	89	90	82	87	¥1.2	256K	Open · Native
14	Keye-VL-2.0Kuaishou	A	86.0	89	86	94	83	¥2	256K	Open · Bolt-on
15	Hunyuan-VisionTencent	A	85.2	88	89	84	80	¥8.6	128K	Closed · Bolt-on
16	InternVL 4.0Shanghai AI Laboratory	A	85.0	96	93	66	78	¥1	256K	Open · Native
17	Step-3StepFun	A-	84.7	88	85	86	85	¥5	256K	Closed · Native
18	Spark X2iFlytek	A-	84.5	86	88	83	79	¥6	128K	Closed · Bolt-on
19	MiniMax M2.5MiniMax	A-	83.8	85	84	82	86	¥8	512K	Closed · Native
20	DeepSeek-OCR 2DeepSeek	B+	81.9	84	99	70	72	¥0.8	64K	Open · Bolt-on
21	Pangu 5.5Huawei	B+	81.6	85	87	80	74	¥20	128K	Closed · Bolt-on
22	Ant Group Ling 2Ant Group	B+	80.8	90	88	60	80	¥5	256K	Closed · Native
23	WITA-Omni PreviewAgiBot	B+	80.6	87	80	92	76	N/A	Undisclosed	Closed · Native
24	Emu3.5BAAI	B+	80.5	86	82	78	70	self-host	128K	Open · Native
25	Zidong Taichu 2.0CASIA	B+	80.1	88	85	84	70	self-host	128K	Open · Native
26	MiniCPM-V 4.5ModelBest	B+	79.6	82	86	74	72	¥0.5	96K	Open · Bolt-on
27	Hikvision Guanlan 3Hikvision	B+	78.2	92	75	91	68	self-host	128K	Closed · Native
28	JD Yanxi 2JD.com	B+	78.0	87	86	55	78	¥6	128K	Closed · Native
29	NetEase Fuxi 2NetEase	B	77.5	86	70	84	75	¥4	128K	Closed · Native
30	NeuroXess-N1NeuroXess	B	77.2	84	82	70	76	self-host	128K	Closed · Native
31	BAGEL-2ByteDance · PKU	B	76.9	84	78	72	68	self-host	64K	Open · Native
32	Cloudwalk Congrong 2Cloudwalk	B	76.1	91	78	68	70	¥7	128K	Closed · Native

Click headers to sort. Price is CNY per 1M output tokens; open models use self-host estimates.

Weights

The composite is not an average

Perception 20%, documents 18%, charts 14%, video 14%, agents 14%, generation 6%, Chinese 6%, ecosystem 8%. If you only care about documents, zero the rest and the ranking changes immediately.

Methodology

20%

Visual perception & grounding

Complex-scene recognition, fine-grained attributes, spatial relations, referring grounding (bbox/point) and Chinese UI elements.

18%

Document parsing & Chinese OCR

Receipts, tables, handwriting, vertical text, formulas and low-quality scans; character accuracy plus layout fidelity.

14%

Chart & document reasoning

Data extraction, trend synthesis, cross-page comparison, unit conversion and multi-step arithmetic beyond OCR.

14%

Video & long-horizon reasoning

Event localization, temporal causality, audio-visual alignment, dense captions and frame consistency.

14%

GUI agents & tool use

Screenshot understanding, click-coordinate accuracy, cross-app long-horizon planning, recovery and multimodal tool orchestration.

Generation & editing

Text-to-image, instruction editing, multi-turn consistency, Chinese glyph rendering and interleaved generation.

Chinese context & culture

Chinese memes, heritage sites, policy documents, dialect speech and local everyday context.

Ecosystem, cost & availability

API stability, open weights, on-prem and domestic-chip support, docs, rate limits and real unit price.

Dimension leaders

Change the dimension, change the winner

The part worth screenshotting: no all-round champion, only scenario champions.

Visual perception & grounding

QMQwen3.8-Max

97/100

Document parsing & Chinese OCR

DSDeepSeek-OCR 2

99/100

Chart & document reasoning

QMQwen3.8-Max

96/100

Video & long-horizon reasoning

豆Doubao Seed 2.0 Pro

96/100

GUI agents & tool use

QMQwen3.8-Max

95/100

Generation & editing

EEmu3.5

98/100

Chinese context & culture

豆Doubao Seed 2.0 Pro

97/100

Ecosystem, cost & availability

豆Doubao Seed 2.0 Pro

98/100

Heatmap

32 models × 8 dimensions

Model	Perception20%	Document18%	Chart QA14%	Video14%	Agent14%	Generation6%	Chinese6%	Ecosystem8%	Overall
Qwen3.8-Max	97	97	96	93	95	90	95	96	95.4
Doubao Seed 2.0 Pro	97	96	94	96	93	88	97	98	95.2
Qwen3.5-397B-A17B	96	97	95	92	92	90	94	98	94.6
GLM-5.3	93	94	91	88	95	85	95	97	92.4
Kimi K3	95	92	93	90	95	84	93	90	92.3
GLM-5	93	95	92	88	91	89	95	92	92.0
MiMo-V2.5	91	90	89	93	90	86	90	88	90.1
SenseNova 6.5 Pro	94	93	90	89	86	87	92	82	89.9
LongCat-Next	92	94	91	87	88	92	89	72	89.2
Qwen3-Omni-30B	90	91	87	90	84	88	91	90	88.9
ERNIE 5.1	90	92	89	85	82	84	93	88	88.1
DeepSeek-V4.1-Flash	90	95	92	78	85	72	86	95	87.9
Qwen3.8-Flash	89	90	86	82	87	80	88	97	87.5
Keye-VL-2.0	89	86	85	94	83	78	86	78	86.0
Hunyuan-Vision	88	89	86	84	80	82	88	80	85.2
InternVL 4.0	96	93	90	66	78	70	89	84	85.0
Step-3	88	85	84	86	85	80	85	78	84.7
Spark X2	86	88	85	83	79	80	90	84	84.5
MiniMax M2.5	85	84	83	82	86	82	84	82	83.8
DeepSeek-OCR 2	84	99	88	70	72	60	90	76	81.9
Pangu 5.5	85	87	84	80	74	70	86	78	81.6
Ant Group Ling 2	90	88	82	60	80	68	92	78	80.8
WITA-Omni Preview	87	80	78	92	76	75	84	60	80.6
Emu3.5	86	82	80	78	70	98	85	70	80.5
Zidong Taichu 2.0	88	85	78	84	70	72	87	65	80.1
MiniCPM-V 4.5	82	86	80	74	72	76	86	80	79.6
Hikvision Guanlan 3	92	75	70	91	68	60	84	70	78.2
JD Yanxi 2	87	86	78	55	78	70	89	76	78.0
NetEase Fuxi 2	86	70	68	84	75	82	88	72	77.5
NeuroXess-N1	84	82	78	70	76	68	88	62	77.2
BAGEL-2	84	78	76	72	68	90	80	70	76.9
Cloudwalk Congrong 2	91	78	72	68	70	65	85	68	76.1

About the ranking and tiers

How are the S/A/B tiers defined?

Composite ≥94 is S+, 92–94 S, 89–92 A+, 85–89 A, 82–85 A-, 78–82 B+, below 78 B. The composite is an eight-dimension weighted score; weights are public on the methodology page. Tiers reflect the composite band only.

Why are GPT and Gemini missing?

This list focuses on native multimodal models built by Chinese teams and legally available in mainland China. Global comparisons live in separate pages and articles to avoid misleading procurement decisions.

Will new models be added?

Yes. A model enters the watchlist with an accessible API or weights, public technical information and at least one verifiable evaluation source. After 30 days it receives a formal score; interim scores are marked provisional.

Why do open models show a price?

That is an estimated self-host cost based on mainstream GPU cloud pricing, throughput and utilisation, for comparison with closed APIs. It is not vendor pricing.

Personal 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.