Seed2.0

围绕大规模生产环境下的使用需求做了系统性优化，旨在帮助突破真实世界中的复杂任务。

API体验链接技术博客Model Card

概述

Seed2.0 系列模型正式发布， 提供 Pro、Lite、Mini 三款不同尺寸的通用 Agent 模型。 该系列通用模型的多模态理解能力实现全面升级，并强化了 LLM 与 Agent 能力，使模型在真实长链路任务中能够稳定推进。 Seed2.0 还进一步把能力边界从竞赛级推理扩展到研究级任务，在高经济价值与科研价值任务评测中达到业界第一梯队水平。

Seed2.0 Pro

侧重长链路推理能力与复杂任务稳定性
适配真实业务中的复杂场景

Seed2.0 Lite

兼顾生成质量与响应速度
适合作为通用生产级模型

Seed2.0 Mini

侧重推理吞吐与部署密度
面向高并发与批量生成场景

模型表现

Seed2.0 在视觉推理与感知上有着显著提升，在 MathVision 等基准测试中达到SOTA水平。 面对动态场景，Seed2.0 强化了对时间序列与运动感知的理解能力，在 MotionBench 等关键测评中处于领先位置。 Seed2.0 还重点强化了指令遵循能力，并在复杂Agent能力评估中，达到业界第一梯队水平。

多模态视觉理解与交互应用

Seed2.0 可以处理复杂视觉输入，并完成实时交互和应用生成。无论是从图像中提取结构化信息，还是通过视觉输入生成交互式内容，Seed2.0 都能高效、稳定地完成任务。

Please enter Chat

专业复杂任务稳定推进

Seed2.0 大幅强化了LLM与Agent表现，在长链路、多步骤指令的任务中稳定且可靠。
Workflow gym - FreeCAD 双凸台建模体积/表面积读取任务

评测结果

Seed2.0 在 LLM、VLM、Agent 领域的任务评估中较 Seed1.8 全面提升，尤其在推理、复杂指令执行和多模态理解方面表现突出。
右滑浏览全部模型结果。
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
Seed2.0 mini
GPT-5.2 High
GPT-5 mini High
Claude Sonnet 4.5
Claude Opus 4.5
Gemini 3 Pro High
Gemini 3 Flash High
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
Seed2.0 mini
GPT-5.2 High
Seed1.8
-
Claude Opus 4.5
Gemini 3 Pro High
-
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
Seed2.0 mini
Seed1.8
Gemini 3 Pro High
Gemini 3 Flash High
Human
-
-
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
GPT-5 mini High
GPT-5.2 High
Claude Sonnet 4.5
Claude Opus 4.5
Gemini 3 Pro High
Gemini 3 Flash High
-
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
Seed2.0 mini
GPT-5.2 High
GPT-5 mini High
Claude Sonnet 4.5
Claude Opus 4.5
Gemini 3 Pro High
Gemini 3 Flash High
Science
MMLU-Pro
87
87.7
83.6
85.9
84.1
88
89.3
90.1
87.8
HLE (no tool, text only)
32.4
28.2
13.3
29.9
17.6
14.5
23.7
33.3
31.7
SimpleQA Verified
36
24
18.9
36.8
26
29.3
48.6
72.1
65.4
HealthBench
57.7
51.2
30
39.4
62.5
28.7
36.3
37.9
51.6
HealthBench - Hard
29.1
20
15.3
36.6
38.6
10.9
11
15
21.5
SuperGPQA
68.7
67.5
61.6
67.9
60.5
65.5
70.6
73.8
72.7
LPFQA
52.6
50.9
47.2
54.4
50.7
54.9
52.6
51.2
51.6
Encyclo-K
65.7
64.5
52.1
61
53
58
63.3
64.9
60
Math
AIME 2026
94.2
88.3
86.7
93.3
92.5
82.5
92.5
93.3
93.3
AIME 2025
98.3
93
87
99
90.3
87
91.3
95
95.2
HMMT Feb 2025
97.3
90
70
100
93.3
79.2
92.9
97.3
100
HMMT Nov 2025
93.3
86.7
80
100
96.7
81.7
93.3
93.3
96.7
MathArenaApex
20.3
4.7
4.2
18.2
2.1
1
1.6
24.5
17.7
MathArenaApex (shortlist)
82.1
52.6
31.1
80.1
43.4
26
47.4
71.4
71.9
BeyondAIME
86.5
76
69
86
72
57
69
83
82
IMOAnswerBench (no tool)
89.3
81.6
71.6
86.6
72.1
60.7
72.6
83.3
84.4
Code
Codeforces
3020
2233
1644
3148
1985
1485
1701
2726
2727
AetherCode
60.6
41.5
29.8
73.8
42.6
16.4
31.6
57.8
56.1
LiveCodeBench (v6)
87.8
81.7
64.1
87.7
62.6
64
84.8
90.7
84.7
STEM
GPQA Diamond
88.9
85.1
79
92.4
82.1
84.3
86.9
91.9
90.7
Superchem (text-only)
51.6
48
16.2
58
34.8
32.4
43.2
63.2
54.4
BABE
50
50.2
40.4
58.1
49.2
44.7
49.3
51.3
55.2
Phybench
74
73
56
74
60
48
69
80
77
FrontierSci-research
25
18.3
3.3
25
18.3
16.7
21.7
15
11.7
FrontierSci-olympiad
74
70
44
75
69
60
71
73
73
General Reasoning
ARC-AGI-1
85.4
75.7
43.3
89.9
54.5
70.9
84
85
86.9
ARC-AGI-2
37.5
14.8
2.3
57.5
3.5
13.6
29.1
31.1
34.3
KORBench
77.5
77
72.8
79.2
74.2
73
77.4
73.9
76
ProcBench
96.6
92.4
80.1
95
87.5
87.5
92.5
90
90
Long Context Performance
MRCR v2 (8-needle)
54
33.6
51.4
89.4
50.1
47.1
56.2
79.7
79
Graphwalks Bfs (<128k)
68.9
82.5
64.1
98
85.5
80.5
92
79.9
84.2
Graphwalks Parents (<128k)
97.6
100
93
99.7
96.6
99
96.2
99.7
99.7
LongBench v2 (128k)
63.8
59.6
52.3
63.2
56.7
62
65
67.4
64
Frames
84.5
83.4
80.5
84
82.9
78.7
84.7
81.9
83.7
DeR2 Bench
58.2
57.3
46.6
69
50.3
58.9
60.4
66.1
66
CL-Bench
20.8
20
14.8
23.9
25.2
18.1
22.6
15.6
16.1
Multilingual
Global PIQA
92.3
92.1
89.2
93.2
91.6
93.9
93.9
95
95.6
MMMLU
88.1
87.7
81.6
90.3
86.3
89.9
91
91.8
91.8
Disco-X
82
80.3
73
76.3
67.7
70.3
78.6
76.8
71.9
Instruction Following
MultiChallenge
68.3
63.2
61.1
59.5
59
57.3
59
68.7
69.3
COLLIE
93.9
94
91.2
96.9
97.4
77.3
79.8
95
96.5
MARS-Bench
85.6
80.5
62.4
87.9
66.1
72.9
87.7
85.6
84.6
Inverse IFEval
78.9
77.1
69.3
72.3
74.8
69.3
72.4
79.6
80.9
Hallucination
LongFact-Objects
92.9
92.2
87
99.2
99.2
98.8
99
98.1
97.9
LongFact-Concepts
92.8
92.4
91.4
99.7
99.5
98.5
98.8
98.5
98.6
FactScore
71.2
62.4
50.4
91.9
96.1
90.6
91.1
92.6
92
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
Seed2.0 mini
GPT-5.2 High
Seed1.8
-
Claude Opus 4.5
Gemini 3 Pro High
-
Multimodal Math
MathVista
89.8
89
85.5
83.1
87.7
-
80.6
89.8
-
MathVision
88.8
86.4
78.1
86.8
81.3
-
74.3
86.1
-
DynaMath
68.9
70.5
58.9
70.1
61.5
-
52.5
63.3
-
MathKangaroo
90.5
86.3
79.8
86.9
73.8
-
69.6
84.4*
-
MathCanvas
61.9
61.1
53.2
55.3
53.6
-
52.9
58.8
-
Multimodal STEM
MMMU
85.4
83.7
79.7
83.7
83.4
-
81.6
87
-
MMMU-Pro
78.2
76
71.4
79.5*
73.2
-
70.8
81.0*
-
EMMA
72
65.5
57
69.4
60.9
-
60.4
66.5
-
SFE
55.6
53.4
48.4
50.1
51.2
-
55.8
61.9
-
HiPhO
74.1
72.5
55.8
77.7
58.3
-
81.8
79.1
-
XLRS-Bench (macro)
54.6
53.7
49.9
49.9
39.9
-
50.4
51.7
-
PhyX (openended)
72.1
62.8
65
71.5
65.9
-
61.3
71
-
Visual Puzzles
LogicVista
81.9
79.6
73.8
81
78.3
-
68.9
80.8
-
VPCT
76
73
48
56
61
-
29
90
-
ZeroBench (main)
12
8
7
11
11
-
4
10
-
ZeroBench (sub)
47.6
42.2
36.2
38.9
37.7
-
30.8
42.2
-
ArcAGI1-Image
88.8
80.9
29.8
93.1
31.4
-
75.8
69.4
-
ArcAGI2-Image
43.3
28.3
1.5
54.4
1.3
-
26.1
21.5
-
VisuLogic
47.4
47.3
40.4
37
35.8
-
27.6
39
-
Perception & Recognition
VLMsAreBiased
77.4
74.8
58.4
28
62
-
21.4
50.6*
-
VLMsAreBlind
98.6
97
93.1
84.2
93
-
77.2
97.5
-
VisFactor
36.8
33.4
23.6
33.6
20.4
-
24.5
45.8
-
RealWorldQA
86
81.7
81.6
82.1
78
-
75.9
84.7
-
BabyVision
60.6
57.5
38.7
37.4
30.2
-
16.2
49.7*
-
General VQA
SimpleVQA
71.4
67.2
68.7
54.1
65.4
-
57.9
69.7
-
HallusionBench
68
66
65.1
67.7
63.9
-
65.3
69.9
-
MME-CC
57
50.2
40.8
44.4
43.4
-
25.2
56.9
-
MMStar
83
80.7
79.1
78.2
79.9
-
73.9
83.1
-
MUIRBench
81.8
76.2
78
77.4
78.7
-
78.9
78.2
-
MTVQA
51.1
51.1
50.6
48.5
47.3
-
53.1
50.8
-
WorldVQA
49.9
44
47.6
26.3
40.4
-
36.6
47.5
-
VibeEval
81.4
76.5
76.5
73.1
74
-
70.3
77.7
-
ViVerBench
75.9
80
73.9
74.8
74.6
-
72.4
75.9
-
Pointing & Counting
CountBench
95.5
97.1
95.5
91.2
96.3
-
90.3
97.3
-
FSC-147↓
11.3
11.9
17.3
21.1
13.6
-
20.9
12.1
-
Point-Bench
81.4
79
77
-
76.5
-
-
85.5*
-
2D & 3D Spatial Understanding
BLINK
79.5
75.6
73.4
70.3
74.3
-
68.1
77.1
-
MMSIBench (circular)
32.5
28.3
19.7
26.1
25.8
-
20.2
25.4
-
TreeBench
64.7
64.2
57.3
58.8
58.5
-
53.6
62.7
-
RefSpatialBench
72.6
66.4
55.6
25.5
56.3
-
-
65.5*
-
DA-2K
92.3
90.3
86.4
78.9
90.7
-
70.3
82.1
-
All-Angles
72.1
65.2
61.3
71.5
61.6
-
63.1
73.5
-
ERQA
68.5
65.8
56.3
59.8
58.8
-
48.3
70.5*
-
Document & Chart Understanding
ChartQAPro
71.2
70.3
65.2
67.6
63
-
-
69
-
OCRBenchv2
62.5
62.4
58.5
55.6
52.6
-
55.5
63.3
-
OmniDocBench 1.5 ↓
0.099
0.102
0.11
0.143*
0.106
-
0.153
0.115*
-
CharXiv-DQ
93.5
93.3
91.9
93.8
88
-
92.7
94.4
-
CharXiv-RQ
80.5
79.9
70.8
82.1*
71.4
-
65.5
81.4*
-
Multimodal LongContext Understanding
DUDE
72.4
72.1
68.9
68.2
69.4
-
55.6
70.1
-
MMLongBench
74.8
70.8
66.7
-
72.4
-
-
73.6
-
LongDocURL
74.7
75.1
71.3
-
74.5
-
-
72
-
MMLongBench-Doc
61.4
55.1
48.9
-
57
-
-
59.5
-
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
Seed2.0 mini
Seed1.8
Gemini 3 Pro High
Gemini 3 Flash High
Human
-
-
Video Knowledge
VideoMMMU
86.9
84.1
80.6
82.7
87.6*
88.1
74.4
-
-
MMVU
78.2
75
69
73.1
76.3
77.9
49.7
-
-
VideoSimpleQA
71.9
66.6
67.7
67.8
71.9
70.7
-
-
-
Video Reasoning
VideoReasonBench
77.8
64.2
40.5
52.8
59.5
61.2
73.8
-
-
Morse-500
37.4
32.2
32.2
29.2
33
32.4
55.4
-
-
VideoHolmes‡
67.4
63.8
58.6
65.5
64.2
65.6
-
-
-
Minerva‡
66.5
63.8
54.7
62.4
65
64.4
-
-
-
Motion & Perception
TVBench
75
71.5
70.5
71.5
71.1
69.6
94.8
-
-
ContPhy
67.4
56.1
55.9
54.9
58
60.5
-
-
-
TempCompass
89.6
87
83.7
86.9
88
88.3
97.3
-
-
EgoTempo
71.8
61.8
67.2
67
65.4
58.4
63.2
-
-
MotionBench
75.2
70.9
64.4
70.6
70.3*
68.9
-
-
-
TOMATO
59.9
57.3
47.4
60.8
55.8
60.8
95.2
-
-
+ Thinking with Tracking
65.3
59.2
51.3
61
-
64
95.2
-
-
Long Video
VideoMME‡
89.5
87.7
81.2
87.8
88.4*
85.2
-
-
-
CGBench
65
59.3
59.2
62.4
65.5
65.3
-
-
-
LongVideoBench
80.3
77.3
74.8
77.4
76.7
74.5
-
-
-
VideoEval-Pro
48
44.3
43.7
45.9
52.7
51.9
-
-
-
LVBench
76.4
73
66.6
73
-
-
-
-
-
Multi Video
CrossVid
60.3
57.7
58.6
57.3
53
48.7
89.2
-
-
Streaming Video
OVBench
69.2
65.5
60.1
65.1
62.7
59.2
-
-
-
LiveSports-3K
78
77.8
73.3
77.5
74.5
71.5
-
-
-
OVOBench
77
76.7
70.4
72.6
70.1
68.7
92.8
-
-
ODVBench
72.5
69.6
65.1
63.5
63.6
56.7
91.4
-
-
ViSpeak
78.5
84
77.5
79
89
86
96
-
-
Capability
Benchmark
Seed2.0 pro
Seed2.0 lite
GPT-5 mini High
GPT-5.2 High
Claude Sonnet 4.5
Claude Opus 4.5
Gemini 3 Pro High
Gemini 3 Flash High
-
Coding Agent
Terminal Bench 2.01
55.8
45
36.9
62.4
45.2
60.2
56.9
60
-
SWE-Lancer
49.4
47.1
43.1
48.9
45.7
56.1
44.3
51.7
-
SWE Bench Verified
76.5
73.5
67.9
80
77.2
80.9
76.2
78
-
Multi-SWE-Bench
45.2
41.1
49.3
47.7
47.7
52.8
50.2
59
-
SWE-Bench Pro
46.9
46
51.7
55.6
48.4
55.4
49.7
46.7
-
SWE Multilingual
71.7
64.4
63.8
68.8
64.1
74
72.7
71.1
-
Scicode
48.5
52.4
40.2
49.7
47.9
52.8
57.7
55
-
SWE-Evo
8.5
10.6
10.4
12.5
16.7
27.1
8.9
12.8
-
Aider Polyglot
80
76
79.1
91.1
82.2
92.4
94.2
92
-
ArtifactsBench
66.6
62.6
68.9
71.1
59.1
68.5
58.4
52.5
-
CodeSimpleQA
58
53.1
57.6
62.3
59.6
63
54.7
53.7
-
SpreadsheetBench Verified
79.1
82.3
58.1
69.9
75.9
78.6
70.8
65.7
-
Search Agent
BrowseComp
77.3
72.1
48.1
77.9（65.3）
43.9（29.5）
67.8（57.2）
59.2
41.5
-
BrowseComp-zh
82.4
82
49.5
76.1
42.4
62.4
66.8
63
-
HLE-text
54.2
49.5
35.8
45.5
32
43.2
46.9
47.6
-
HLE-Verified
73.6
70.7
56.4
68.5
37.6
56.6
67.5
71.8
-
WideSearch
74.7
74.5
37.7
76.8
65.1
76.2（71.7）
67.3
64
-
FinSearchComp
70.2
65.1
38.1
73.8
58.6
66.2
52.7
54.8
-
DeepSearchQA
77.4
67.7
16.7
71.3（66.4）
36.3
（76.1）41.6
63.9
54.7
-
Seal-0
49.5
52.3
34.2
51.4
53.4
47.7
45.5
37.8
-
Tool Use
τ2-Bench (retail)
90.4
90.9
-
82
86.2
88.9
85.3
88.6
-
τ2-Bench (telecom)
94.2
92.1
-
98.7
98
98.2
98
94.7
-
MCP-Mark
54.7
46.7
30.2
57.5
32.1
42.3
53.9
40.5
-
BFCL-v4
73.4
72.9
57.9
65.9
72.9
76.5
71
65
-
VitaBench
47
41.8
20.8
41.8
40.8
55.3
48.8
46.7
-
Deep Research
DeepConsult
61.1
60.3
49.8
54.3
55.8
61
48
26
-
DeepResearchBench
53.3
54.4
50.7
52.2
47.2
50.6
49.6
46.1
-
ResearchRubrics
50.7
50.8
43.6
42.3
38.6
45
37.7
36.9
-
Vision Agent
Minedojo Verified
49
39.7
20.3
18.3
-
-
23.3
24.3
-
MM-BrowseComp
48.8
45.1
17.9
26.3
-
-
25
22.8
-
HLE-VL
39.2
35.8
18.7
31
-
-
36
36.8
-
Science Discovery
Scicode
52.1
52.4
40.2
49.7
47.9
52.8
57.7
55
-
FrontierSci-research
23.3
18.3
18.3
25
16.7
21.7
15
11.7
-
Superchem (text-only)
53
48
34.8
58
32.4
43.2
63.2
54.4
-
BIObench
53.5
50.2
49.2
58.1
44.7
49.3
51.3
55.2
-
AInstein Bench
47.7
38.3
35
41.3
33.7
44
42.8
44
-
Vibe Coding
NL2Repo-Bench
27.9
24.6
19.5
49.3
39.9
43.2
34.2
27.6
-
NL2Repo (Pass@1)
3
1
2
8
3
3
4
2
-
ArtifactsBench
66.6
62.6
68.9
71.1
59.1
68.5
58.4
52.5
-
SWE-Bench Pro
46.9
46
51.7
55.6
48.4
55.4
49.7
46.7
-
Terminal Bench 2.01
55.8
45
36.9
62.4
45.2
60.2
54.2
60
-
Context Learning
KORBench
77.2
77
74.2
79.2
73
77.4
73.9
76
-
DeR2 Bench
58.2
57.3
50.3
69
58.9
60.4
66.1
66
-
CL-Bench
21.5
20
25.2
23.9
18.1
22.6
15.6
16.1
-
ToB-Complex Workflows
64.7
68.2
53.4
45
61
64.8
69.2
64.9
-
ToB-Reference Q&A
72.4
62
42.5
63.6
58.9
67.9
68.3
61.3
-
Real World Tasks
HealthBench - Hard
28.3
20
38.6
36.6
10.9
11
15
21.5
-
GDPVal-Diamond
21.3
23.2
11.5
26.9
15.2
20.7
19.4
6.46
-
XPert Bench
64.5
63.3
47.6
53.3
44.7
50.5
53.1
50.1
-
ToB-K12 Education
62.8
63.8
53.7
61.6
50.1
56.2
59.4
59
-
ToB-Compositional Tasks
59.1
54.8
47.5
51.5
57.3
63.6
64.8
57.7
-
ToB-Text Classification
69
64.5
52.7
62.1
64.5
63.9
67.5
61.9
-
ToB-Information Extraction
52
48.4
40.5
44.7
48.3
50.1
49
45.4
-
World Travel (VLM)
12
14
2.7
19.33
2.67
14
8
7.3
-
World Travel (TEXT)
23.3
24
15.3
32.67
10
21.3
14.7
13.3
-
带有 * 的结果引自技术报告。
带有 ‡ 的基准测试在评估时包含了字幕内容。
1 ：使用 Terminus2。