Seed2.0

Seed2.0 has been systematically optimized to meet the requirements of large-scale production deployment, and is designed to help tackle complex real-world tasks.

Model CardTech BlogTry NowAPI

Overview

The Seed2.0 series model has been officially released, offering three general-purpose agent models of varying sizes—Pro, Lite, and Mini. The general-purpose models in this series deliver a comprehensive upgrade in multimodal understanding, with strengthened LLM and Agent capabilities that enable steady progression in real-world long-horizon tasks. Seed2.0 further expands its capability frontier from competition-level reasoning to research-grade tasks, achieving first-tier industry performance in evaluations on high economic-value and high scientific-value workloads.

*Seed2.0 Lite was upgraded at the end of April. As the first omni-modal understanding model in the Seed foundation model series, it natively supports unified understanding across video, image, audio, and text, while also featuring upgraded Agent, Coding, and GUI capabilities.

Seed2.0 Pro

Focuses on long-chain reasoning and robustness in complex workflows. Optimized for complex scenarios in real-world tasks.

Seed2.0 Lite

Balances output quality and response speed.
Ideal as a general-purpose, production-grade model.

Seed2.0 Mini

Optimized for inference throughput and deployment density. Designed for high concurrency and batch generation scenarios.

Model Performance

Seed2.0 delivers significant enhancements in visual reasoning and perception, and achieves SOTA performance on benchmarks, such as BabyVision. In dynamic video scenarios, Seed2.0 strengthens its foundational capabilities in temporal understanding and motion perception, and achieves SOTA results on multiple video reasoning benchmarks. Seed2.0 further enhances its instruction-following capabilities and achieves top-tier industry performance when evaluated on complex Agent capabilities. *The latest version of Seed2.0 Lite enables cross-modal reasoning that combines audio and visual information, achieving industry-leading performance on both video and audio understanding benchmarks. Beyond standard benchmarks, we place greater emphasis on actual user experience. Before this version was released, we invited over 50 Ark users to evaluate the Coding Agent capabilities of the latest Seed2.0 Lite model. The results show that the new version also delivers a clear improvement in this area.

Omni-Modal Understanding and Interactive Applications

Seed2.0 can process complex visual inputs and enable real-time interaction and app generation. Whether extracting structured information from images or generating interactive content via visual inputs, Seed2.0 handles tasks fast and reliably. Additionally, with its late-April update, Seed2.0 Lite now supports audio input, enabling omni-modal understanding.

Steady progression on sophisticated professional tasks

Seed2.0 significantly enhances the performance of its LLM and Agent, maintaining high stability and reliability when executing long-horizon, multi-step instructions.
Workflow Gym – FreeCAD Double Boss Modeling: Volume and Surface Area Extraction Task

Evaluation Results

We conducted a comprehensive evaluation of the Seed2.0 series, which demonstrated strong performance across key tasks such as reasoning, complex instruction following, and multimodal understanding. The latest Seed2.0 Lite model shows significant performance improvements and achieves SOTA results on multiple video and audio understanding benchmarks.
*Swipe right to view all model evaluation results.
Capability
Benchmark
Seed2.0 Lite（0428）
Seed2.0 Lite（0215）
Seed2.0 Pro（0215）
GPT-5.4 Mini
Gemini 3 Flash
-
-
Capability
Benchmark
Seed2.0 Lite（0428）
Seed2.0 Lite（0215）
Seed2.0 Pro（0215）
GPT-5.4 High
Gemini 3 Flash
Gemini 3.1 Pro High
-
Capability
Benchmark
Seed2.0 Lite（0428）
Seed1.8
Claude Opus 4.7
Claude Sonnet 4.6
Claude Sonnet 4.5
GPT-5.4 High
Gemini 3.1 Pro
Capability
Benchmark
Seed2.0 Lite（0428）
Seed2.0 Lite（0215）
Seed2.0 Pro（0215）
Seed2.0 Mini（0215）
Gemini 3 Pro High
Gemini 3 Flash High
-
Capability
Benchmark
Seed2.0 Lite（0428）
Gemini-3.1-Pro
-
-
-
-
-
Capability
Benchmark
Seed2.0 Lite（0428）
Seed2.0 Lite（0215）
Seed2.0 Pro（0215）
GPT-5.4 Mini
Gemini 3 Flash
-
-
Knowledge
GPQA Diamond
88.4%
85.1%
88.9%
88.0%
90.7%
-
-
SuperGPQA
69.6%
67.5%
68.7%
63.9%
72.7%
-
-
HLE (no tool, text only)
25.7%
28.2%
32.4%
28.2%
31.7%
-
-
Reasoning
BeyondAIME
79.0%
76.0%
86.5%
80.0%
82.0%
-
-
FrontierSci-olympiad
72.0%
70.0%
74.0%
70.0%
73.0%
-
-
Superchem (text-only)
55.0%
48.0%
51.6%
29.1%
54.4%
-
-
BABE
57.9%
50.2%
53.5%
49.0%
55.2%
-
-
Instruction Following
CL-Bench
20.1%
20.0%
20.8%
14.9%
16.1%
-
-
MultiChallenge
69.9%
63.2%
68.3%
62.5%
69.3%
-
-
SearchAgent
WideSearch
70.3%
74.5%
74.7%
73.0%
64.0%
-
-
BrowseComp
64.0%
72.1%
77.3%
61.3%
41.5%
-
-
ResearchRubrics
59.2%
50.8%
50.7%
47.1%
36.9%
-
-
XPert Bench
56.8%
63.3%
64.5%
41.8%
50.1%
-
-
Real World
‌SkillsBench
43.7%
42.1%
42.3%
45.4%
26.4%
-
-
GDPval
53.1%
47.3%
54.4%
50.6%
13.7%
-
-
FinSearchComp
63.8%
65.1%
70.2%
61.8%
43.7%
-
-
Tob-Agent
51.4%
45.2%
52.6%
43.0%
37.4%
-
-
CodingAgent
SWE Multilingual
66.6%
64.4%
71.7%
73.6%
71.1%
-
-
SWE-Bench Pro
46.6%
46.0%
46.9%
54.4%
46.7%
-
-
NL2Repo-Bench
28.7%
24.6%
27.9%
37.3%
27.6%
-
-
PaperBench
52.5%
54.6%
53.8%
49.1%
33.9%
-
-
Terminal Bench 2.0
43.3%
45.0%
55.8%
60.0%
60.0%
-
-
Vibe Coding 人工评估
49.4%
48.7%
48.4%
57.4%
56.9%
-
-
Capability
Benchmark
Seed2.0 Lite（0428）
Seed2.0 Lite（0215）
Seed2.0 Pro（0215）
GPT-5.4 High
Gemini 3 Flash
Gemini 3.1 Pro High
-
STEM
MathVision
89.8
86.4
88.8
90.6
87.5
89.0
-
MMMU_Pro
78.4
76.0
78.2
79.2
80.4
82.5
-
HiPhO
83.8
72.5
74.1
84.3
78.0
86.6
-
MedXpertQA-MM
79.6
64.0
68.1
76.9
78.0
80.2
-
Perception
BabyVision
64.7
57.5
60.6
53.4
47.2
54.4
-
VLMBias
80.6
74.8
77.4
42.8
66.1
73.5
-
Visual Knowledge
SimpleVQA
72.7
67.2
71.4
56.0
68.4
70.5
-
WorldVQA
50.2
44.0
49.9
30.2
46.5
44.4
-
InfoGraphics
CharXiv-DQ
94.5
93.3
93.5
94.1
94.0
94.9
-
CharXiv-RQ
82.4
79.9
80.5
82.6
79.7
84.0
-
Embodied
ERQA
71.5
65.8
68.5
64.5
65.8
70.8
-
Capability
Benchmark
Seed2.0 Lite（0428）
Seed1.8
Claude Opus 4.7
Claude Sonnet 4.6
Claude Sonnet 4.5
GPT-5.4 High
Gemini 3.1 Pro
GUI
OSWorld-Verfied
64.4%
61.9%
78.0%
72.5%
62.9%
75.0%
64.0%
MobileWorld
64.6%
52.1%
56.4%
-
47.8%
-
57.3%
Capability
Benchmark
Seed2.0 Lite（0428）
Seed2.0 Lite（0215）
Seed2.0 Pro（0215）
Seed2.0 Mini（0215）
Gemini 3 Pro High
Gemini 3 Flash High
-
Video Knowledge
VideoMMMU
88.3
84.1
86.9
80.6
87.6*
88.1
-
MMVU
76.7
75.0
78.2
69.0
76.3
77.9
-
VideoSimpleQA-v2
69.0
65.0
71.5
64.9
-
-
-
VideoSimpleQA
71.7
66.6
71.9
67.7
72.4
70.0
-
SciVideo
70.3
51.4
52.3
35.3
-
74.1
-
Video Reasoning
VideoReasonBench
59.4
64.2
77.8
40.5
59.5
61.2
-
VideoHolmes
67.4
63.8
67.4
58.6
64.2
65.6
-
Minerva
68.5
63.8
66.5
54.7
65.0
64.4
-
Motion & Perception
TVBench
80.4
71.5
75.0
70.5
71.1
69.6
-
TOMATO
72.5
57.3
59.9
47.4
55.8
60.8
-
EgoTempo
68.4
61.8
71.8
67.2
65.4
58.4
-
MotionBench
72.4
70.9
75.2
65.1
70.3
68.9
-
ContPhy
62.4
56.1
67.4
55.9
61.1
62.0
-
Morese-500
34.6
32.2
37.4
32.2
33.0
32.4
-
Long Video
VideoMME
89.0
87.7
89.5
81.2
88.4*
85.2
-
VideoMMEv2
64.9
-
60.5*
-
66.1*
61.1*
-
CGBench
65.5
59.3
65.0
-
65.5
65.3
-
LongVideoBench
79.0
77.3
80.3
74.8
78.2
74.5
-
LVBench
76.4
73.0
76.4
66.6
-
-
-
VideoEval-Pro
49.5
44.3
47.3
43.7
-
51.9
-
Streaming Video
OVBench
63.2
65.5
69.2
60.1
63.5
59.2
-
ODVBench
66.0
69.6
72.5
65.1
63.6
56.7
-
LiveSports-3K
78.1
77.8
78.0
73.3
74.5
73.2
-
OVOBench
75.4
76.7
77.0
70.4
70.1
68.7
-
ViSpeak
87.0
84.0
78.5
77.5
89.0
88.0
-
Multi-video
CrossVid
63.7
57.7
61.0
58.6
53.0
48.7
-
Visual-Audio Understanding
OmniVideoBench
61.7
44.5
49.5
40.8
61.4(3.1Pro)
-
-
AVMeme
69.5
60.6
61.2
50.7
77.3(3.1Pro)
-
-
JointAVBench
69.5
56.7
62.3
52.7
-
-
-
WorldSense
67.3
57.0
57.0
52.7
65.5(3.1Pro)
-
-
Capability
Benchmark
Seed2.0 Lite（0428）
Gemini-3.1-Pro
-
-
-
-
-
Audio Understanding
MMSU
86.54
85.94
-
-
-
-
-
WildSpeech
75.81
75.41
-
-
-
-
-
ASR
WenetSpeech test-net
4.47
9.52
-
-
-
-
-
WenetSpeech test-meeting
5.31
12.80
-
-
-
-
-
Librispeech test-clean
1.07
1.94
-
-
-
-
-
Librispeech test-other
2.17
3.60
-
-
-
-
-
S2TT
Fleurs(15 langs)(zh/en<->xx)
74.70
73.14
-
-
-
-
-
ASR performance is evaluated using WER/CER, where lower values indicate better performance.