Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity
Seed2.0 is a multimodal model series optimized for real-world agentic workflows, balancing frontier-level reasoning with production-grade cost efficiency.
How does the Seed2.0 model family perform across fundamental language, vision, video, and complex agentic tasks in real-world enterprise environments?
Large Language Models (LLMs) often excel at competition-level reasoning but fail to reliably complete end-to-end, multi-step tasks in professional environments. This gap stems from a lack of long-horizon planning and insufficient ingestion of domain-specific, long-tail knowledge. Seed2.0 addresses this by prioritizing multimodal understanding, long-context reasoning, and structured generation within a workflow-oriented foundation. The series is designed to integrate directly into enterprise digital infrastructure rather than serving as a peripheral tool. The model series achieves performance parity with international frontier models on core benchmarks while offering token pricing approximately an order of magnitude lower than comparable high-end alternatives.
Paper Primer
Seed2.0 functions as a workflow-oriented foundation, emphasizing the ability to synthesize heterogeneous data and execute multi-step instructions. The core move is a systematic ingestion of long-tail professional knowledge combined with architectural optimizations that allow for tiered deployment (Pro, Lite, Mini) to match specific latency and cost requirements.
Seed2.0 Pro achieves gold-medal level performance on 2025 Olympiad-level mathematical competitions.
The model scored 35/42 on the IMO 2025 and 114/126 on the CMO 2025, meeting the gold medal threshold for both.
Seed2.0 provides a significant cost advantage for enterprise-scale deployments.
Seed2.0 Pro input/output pricing is $0.47/\$2.37 per million tokens, compared to $5.00/\$25.00 for Claude-Opus-4.5-thinking.
Why does the paper emphasize "long-tail" professional knowledge over general factual recall?
Real-world enterprise workflows rely on domain-specific information—such as internal technical manuals or specialized legal texts—that generic models often underserve, making mastery of this long-tail data essential for reliable end-to-end task completion.
What is the primary difference between Seed2.0 and previous "conversational" models?
Seed2.0 is designed as a workflow-oriented foundation that integrates into production systems to process heterogeneous data and execute multi-step instructions, rather than acting as a lightweight assistant for isolated prompts.
Introduction to Seed2.0
Introduces the Seed2.0 model family and the real‑world complexity gap it aims to close.
Large language models now underpin modern digital infrastructure, serving both professional and personal applications at massive scale. The Seed team’s prior model series (Seed1.x) already powers hundreds of millions of daily active users.
Seed2.0 is a family of models that scale from basic language reasoning up to sophisticated, real‑world agentic tasks in enterprise settings.
Seed2.0 targets four user‑experience pillars: robust visual and multimodal understanding, low inference latency, reliable complex instruction execution, and seamless coding assistance. These pillars guide the design of three model sizes—Pro, Lite, and Mini—so developers can trade speed for capability.
Current agentic systems excel at competition‑level problems yet often fail to complete practical, end‑to‑end tasks such as building a well‑designed application in one pass. Two root causes are long‑horizon workflow construction and the scarcity of domain‑specific long‑tail knowledge.
To measure progress, we define an evaluation framework spanning Science Discovery, Vibe Coding, Context Learning, and Real‑World Tasks, each probing a core aspect of complex, long‑horizon agent performance.
The Seed2.0 family provides the scalable, low‑latency backbone that modern digital services need to embed sophisticated, real‑world agentic capabilities.
Deployment Patterns and Developer Behavior
Seed2.0’s pricing is an order of magnitude lower, enabling cost‑effective enterprise MaaS deployments.
Seed2.0 pricing is roughly ten times lower than leading foundation models.
Table 1 shows the decode price for Seed2.0 Mini at \$0.31 per 1 M tokens versus \$14.00 for GPT‑5.2 High.
MaaS (Model-as-a-Service) delivers AI models as hosted APIs that enterprises embed directly into production workflows, treating the model as a reusable service rather than a local component.
**Figure 1.** MaaS usage distribution in mainland China. Left: Industry traffic distribution showing strong dominance of the Internet sector. Right: Business Customer Usage Scenario Distribution. This data statistics is named “Doubao Collaboration Incentive Program”, which sourced from the authorization of customers who have signed the Data Authorization Agreement.
**Figure 2.** Query distribution by development and requirement domain. Frontend development and bug fixing substantially dominate agentic coding requests, each far exceeding their respective category alternatives.
Evaluation Framework and Methodology
We outline the evaluation framework that quantifies Seed2.0’s language, vision, and agentic abilities.
The evaluation framework is built to expose concrete strengths and gaps of Seed2.0 across language, vision, and agentic behaviours.
It measures how reliably a model can reason, obey complex instructions, and retain broad factual knowledge across domains.
Question 1: model outputs “4” → exact match, counted as correct.
Question 2: model outputs “Paris” → exact match, counted as correct.
Question 3: model outputs “Water boils at 100 °C because of atmospheric pressure” → missing the pressure justification, counted as incorrect.
Accuracy = 2 / 3 ≈ 66.7 %.
This toy calculation shows how the capacity metric collapses diverse abilities into a single accuracy figure, highlighting that a model can excel on factual recall yet still miss nuanced scientific reasoning.
How does Fundamental Language Capacity differ from a standard language‑model perplexity score?
Perplexity only measures next‑token prediction on a held‑out corpus; Fundamental Language Capacity aggregates performance on reasoning, instruction following, and domain‑specific knowledge tasks, each requiring multi‑step inference or factual synthesis that perplexity does not capture.
Beyond language, the framework also quantifies vision abilities, agentic planning, and end‑to‑end task execution.
The image contains a table summarizing performance in two competitions, IMO 2025 and CMO 2025, across six problems (P1-P6), showing individual scores, overall totals, and medal results.
This comprehensive suite lets practitioners pinpoint exactly where Seed2.0 excels or falls short, guiding both deployment decisions and future research directions.
Fundamental Language Performance
Seed2.0 Pro ranks among the top models on core language benchmarks.
Seed2.0 Pro outperforms Seed‑1.8 on the in‑house Chinese benchmark, achieving 75.26 % overall—a +2.37 % absolute gain.
Table 7 shows Seed2.0 Pro leading in nine of ten categories, with the highest overall score of 75.26 % versus 72.89 % for Seed‑1.8.
Across the large‑model suite (Table 3), Seed2.0 Pro sits comfortably within the international leading group, matching or surpassing GPT‑5.2 and Gemini‑3‑Pro on most categories.
On the efficient‑model suite (Table 4), Seed2.0 Lite and Mini deliver strong efficiency‑quality trade‑offs, often trailing only Gemini‑3‑Flash while remaining far ahead of GPT‑5‑mini.
Seed2.0 Pro leads on HealthBench (57.7) and HealthBench‑Hard (29.1), far exceeding the nearest competitor’s 37.9 and 15.0 respectively, which matters for production queries that probe obscure domain knowledge.
The model also attains gold‑medal level performance on the 2025 International Mathematical Olympiad and China Mathematical Olympiad, confirming its advanced mathematical reasoning capability.
In code generation, Seed2.0 Pro reaches a Codeforces Elo of 3020, surpassing all listed baselines and demonstrating competitive front‑end generation quality.
**Table 3.** Evaluation on Fundamental Language Capacity Benchmarks (Large Models). The highest score is marked in bold, and the second is underlined.
**Table 4.** Evaluation on Fundamental Language Capacity Benchmarks (Efficient Models). The highest score is marked in bold, and the second is underlined.
**Table.** Performance comparison across different benchmarks and model versions.
**Table 6.** Complex instruction-following benchmark across multiple test sets and instruction categories.
**Table 7.** Complex instruction following performance on in-house Chinese benchmark. Bold indicates the better result.
Seed2.0 Pro consistently ranks at the top of standard language benchmarks, confirming its readiness for enterprise‑grade deployments.
Vision Task Performance
Seed2.0 Pro leads visual‑language benchmarks, topping most categories.
Seed2.0 Pro leads in 9 of 9 capability categories on the visual‑language suite.
Table 8 shows Seed2.0 Pro attaining the highest Pass@1 score in the majority of benchmarks across all categories.
Mathematical reasoning: Seed2.0 Pro scores 88.8 on MathVision, 90.5 on MathKangaroo, 61.9 on MathCanvas, and ties at 89.8 on MathVista; Seed2.0 Lite tops DynaMath with 70.5.
STEM tasks: Seed2.0 Pro leads EMMA (72.0), XLRS‑Bench (54.6) and PhyX (72.1), while improving HiPhO by +15.8 points and MMMU‑Pro by +5.0 over Seed1.8.
Visual puzzles & logic: Seed2.0 Pro attains 81.4 on LogicVista, 12.0 (main) and 47.6 (sub) on ZeroBench, and 47.4 on VisuLogic, while GPT‑5.2 leads ArcAGI‑Image; Seed2.0 Pro follows with 88.8 (ArcAGI1) and 43.3 (ArcAGI2).
Perception & recognition: Seed2.0 Pro sets new records on VLMsAreBiased (77.4), VLMsAreBlind (98.6) and BabyVision (60.6).
General VQA: Seed2.0 Pro leads SimpleVQA (71.4), MUIRBench (81.8) and VibeEval (81.4), outperforming Seed1.8 across the board.
Spatial understanding & counting: Seed2.0 Pro achieves state‑of‑the‑art on DA‑2K (92.3), RefSpatialBench (72.6) and BLINK (79.5); its counting error on FSC‑147 drops to 11.3 MAE.
Document & long‑context understanding: Seed2.0 Pro dominates DUDE (72.4), MMLongBench (74.8), MMLongBench‑Doc (61.4), ChartQAPro (71.2) and leads OmniDocBench 1.5.
**Table 8.** Performance of Seed2.0 on public visual-language benchmarks compared to previous models. We report Pass@1 in these benchmarks. The best score for each benchmark is marked in **bold**, and the second best is <u>underlined</u>. Results marked with an * are sourced from the technical report.
Video Task Performance
Seed2.0 Pro sets new records on long‑video benchmarks.
Seed2.0 Pro achieves the highest score of 89.5 on the VideoMME long‑video benchmark, outpacing all prior models.
Table 9 lists Seed2.0 Pro at 89.5, while the next best (Gemini‑3‑Pro) scores lower.
Beyond the headline scores, Seed2.0 excels across knowledge (MMVU, VideoMMMU, VideoSimpleQA), reasoning (surpassing human on VideoReasonBench), motion perception (TOMATO gains), multi‑video understanding (60.3 on CrossVid), and streaming scenarios where the Lite variant outperforms Gemini‑3‑Pro/Flash.
**Table 9.** Performance of Seed2.0 on public video understanding benchmarks compared to previous models. The highest score in each benchmark is marked in bold, and the second is underlined. For benchmarks marked with a ‡, we include subtitles for evaluation. Results marked with an * are sourced from the technical report.
**Table.** Benchmark Representative Findings and Efficiency Profile
Agentic Capacity Evaluation
Seed2.0’s agentic performance across search, research, vision, coding, and tool‑use benchmarks.
Agentic Capacity measures how well a model can act as an autonomous tool‑using assistant across diverse tasks such as searching, reasoning, coding, and multimodal interaction.
Seed2.0 Pro attains the highest score of 90.4 on the $\tau$ 2‑Bench (retail) benchmark.
Table 11 shows Seed2.0 Pro scoring 90.4, surpassing GPT‑5.2 (86.2) and Claude‑Opus‑4.5 (88.9).
**Table 11.** Evaluation on Fundamental Agentic Capacity Benchmarks (Large Models). The highest score is marked in bold, and the second is underlined. Some scores differ greatly from the evaluation results in the tech reports by other organizations. The scores in parentheses represent the results under the aligned settings then.
**Table 12.** Evaluation on Fundamental Agentic Capacity Benchmarks (Efficient Models). The highest score is marked in bold, and the second is underlined.
Scientific and Economic Task Performance
Seed2.0 excels on high‑value scientific and economic benchmarks, delivering top‑tier performance.
Seed2.0 Pro attains a 73.02 % Pass@8 score on ICPC competition benchmarks, surpassing GPT‑5.2 (65.08 %) and Gemini‑3‑Pro (63.49 %).
Pass@8 scores are computed over five official ICPC contests; Seed2.0 Pro solves the highest proportion of problems within eight generation attempts.
It quantifies how well a model translates raw capability into concrete value for research‑driven and revenue‑driven workloads.
How does Seed2.0’s performance on scientific discovery differ from its performance on economic‑value tasks?
On scientific discovery benchmarks Seed2.0 Pro leads with the highest raw scores, reflecting deep reasoning and code synthesis for research‑style problems. On economic‑value tasks it remains competitive but sometimes trails specialized baselines that are tuned for short‑turn instruction following, such as GPT‑5.2 on Vibe Coding.
**Table 13.** Evaluation on Advanced Economically and Scientifically Valuable Tasks (Large Models). The highest score is marked in bold, and the second is underlined.
**Table 14.** Evaluation on Advanced Economically and Scientifically Valuable Tasks (Efficient Models). The highest score is marked in bold, and the second is underlined.
**Figure 4.** Case study of Seed2.0 performing the Project Repository Construction Task on NL2Repo [33].
**Figure 5.** Case study of Seed2.0 performing the Iterative Code Debugging task on SweBenchPro [31]
**Figure 6.** The case of Seed2.0 with Complex Iterative Debugging Task on SWE-EVO.
Seed2.0’s strongest advantage lies in high‑value scientific discovery, where its reasoning and code synthesis outperform competing models.
Questions & answers
What is Seed2.0 and what is its main contribution?
Seed2.0 is a workflow-oriented foundation model series from the Seed team, offered in three sizes (Pro, Lite, Mini), that prioritizes multimodal understanding, long-context reasoning, and structured generation for enterprise use. Its main contribution is matching international frontier model performance on core benchmarks while offering token pricing approximately an order of magnitude lower than comparable high-end alternatives.
What problem does Seed2.0 address?
Seed2.0 addresses the gap between LLMs excelling at competition-level reasoning and their failure to reliably complete end-to-end, multi-step tasks in professional environments. The paper identifies two root causes: insufficient long-horizon planning and a lack of domain-specific, long-tail knowledge ingestion.
Why does Seed2.0 emphasize long-tail professional knowledge?
Real-world enterprise workflows rely on domain-specific information—such as internal technical manuals or specialized legal texts—that generic models often underserve, making mastery of this long-tail data essential for reliable end-to-end task completion.
How does Seed2.0 differ from conversational or assistant-style models?
Seed2.0 is designed as a workflow-oriented foundation that integrates into production systems to process heterogeneous data and execute multi-step instructions, rather than acting as a lightweight assistant for isolated prompts.
What are the four pillars of the Seed2.0 user-experience design?
The four pillars are robust visual and multimodal understanding, low inference latency, reliable complex instruction execution, and seamless coding assistance. These pillars guide the design of the three model sizes—Pro, Lite, and Mini—so developers can trade speed for capability.
What evaluation framework does the paper use to assess Seed2.0?
The paper defines an evaluation framework spanning four dimensions: Science Discovery, Vibe Coding, Context Learning, and Real-World Tasks, each probing a core aspect of complex, long-horizon agent performance. It also includes Fundamental Language Capacity, vision benchmarks, and agentic planning assessments.
How does Fundamental Language Capacity differ from a standard perplexity score?
Perplexity only measures next-token prediction on a held-out corpus, whereas Fundamental Language Capacity aggregates performance on reasoning, instruction following, and domain-specific knowledge tasks, each requiring multi-step inference or factual synthesis that perplexity does not capture.
What are Seed2.0 Pro's key results on language benchmarks?
Seed2.0 Pro sits within the international leading group, matching or surpassing GPT-5.2 and Gemini-3-Pro on most categories. It leads on HealthBench (57.7) and HealthBench-Hard (29.1), far exceeding the nearest competitor's scores of 37.9 and 15.0 respectively, and achieves a Codeforces Elo of 3020 in code generation.
How does Seed2.0 perform on mathematical reasoning?
Seed2.0 Pro attains gold-medal level performance on the 2025 International Mathematical Olympiad and China Mathematical Olympiad, and scores 88.8 on MathVision, 90.5 on MathKangaroo, 61.9 on MathCanvas, and 89.8 on MathVista. Seed2.0 Lite tops DynaMath with 70.5.
What are Seed2.0 Pro's notable vision and multimodal benchmark results?
Seed2.0 Pro leads EMMA (72.0), XLRS-Bench (54.6), PhyX (72.1), LogicVista (81.4), SimpleVQA (71.4), MUIRBench (81.8), DA-2K (92.3), DUDE (72.4), MMLongBench (74.8), and sets new records on VLMsAreBlind (98.6) and BabyVision (60.6), among others. It also improves HiPhO by +15.8 points and MMMU-Pro by +5.0 over Seed1.8.
How does Seed2.0 perform on video understanding tasks?
Seed2.0 excels across knowledge benchmarks (MMVU, VideoMMMU, VideoSimpleQA), surpasses human performance on VideoReasonBench, achieves 60.3 on CrossVid for multi-video understanding, and the Lite variant outperforms Gemini-3-Pro and Gemini-3-Flash in streaming scenarios.
How does Seed2.0's performance on scientific discovery compare to economic-value tasks?
On scientific discovery benchmarks, Seed2.0 Pro leads with the highest raw scores, reflecting deep reasoning and code synthesis for research-style problems. On economic-value tasks it remains competitive but sometimes trails specialized baselines tuned for short-turn instruction following, such as GPT-5.2 on Vibe Coding.
How do Seed2.0 Lite and Mini compare to other efficient models?
On the efficient-model suite, Seed2.0 Lite and Mini deliver strong efficiency-quality trade-offs, often trailing only Gemini-3-Flash while remaining far ahead of GPT-5-mini.
What is the pricing advantage of Seed2.0 relative to frontier alternatives?
The paper states that Seed2.0 achieves performance parity with international frontier models while offering token pricing approximately an order of magnitude lower than comparable high-end alternatives. The paper does not specify exact per-token prices.
What prior work does Seed2.0 build upon?
Seed2.0 is developed by the Seed team, whose prior model series Seed1.x already powers hundreds of millions of daily active users. The paper references improvements over Seed1.8 on specific benchmarks such as HiPhO (+15.8 points) and MMMU-Pro (+5.0).
What are the limitations or gaps acknowledged by the paper?
The paper acknowledges that Seed2.0 sometimes trails specialized baselines on economic-value and short-turn instruction-following tasks such as Vibe Coding, where GPT-5.2 leads. It also notes that GPT-5.2 leads on ArcAGI-Image among visual puzzle benchmarks.
Who developed Seed2.0 and when was the paper published?
Seed2.0 was developed by the Seed team. The paper is available on arXiv (arxiv.org/abs/2607.00248); the paper does not specify individual author names or a formal venue beyond the arXiv preprint.
How is Seed2.0 intended to be deployed in practice?
Seed2.0 is designed to integrate directly into enterprise digital infrastructure through a tiered deployment model—Pro, Lite, and Mini—allowing developers to select the appropriate size based on latency and cost requirements. The paper does not provide specific API or reproducibility details.
Key terms
- Seed2.0
- A tiered multimodal large language model series (Pro, Lite, Mini) developed by the Seed team, designed for enterprise workflow integration with an emphasis on long-context reasoning and domain-specific knowledge.
- long-tail knowledge
- Rare, domain-specific information—such as specialized legal texts or internal technical manuals—that appears infrequently in general training data but is critical for professional enterprise tasks.
- workflow-oriented foundation
- A model design philosophy that prioritizes integration into multi-step production pipelines and heterogeneous data processing, rather than serving as a standalone conversational assistant.
- long-horizon planning
- The ability of a model to construct and execute a sequence of interdependent steps toward a complex goal without losing coherence or context over many turns or actions.
- Fundamental Language Capacity
- An evaluation metric that aggregates performance across reasoning, instruction following, and domain-specific knowledge tasks, going beyond simple next-token prediction accuracy.
- agentic capacity
- A model's ability to autonomously plan, make decisions, and execute multi-step tasks in dynamic environments, as opposed to responding to single isolated prompts.
- HealthBench
- A benchmark designed to evaluate a model's performance on health-related queries, including obscure domain knowledge, with a harder subset called HealthBench-Hard.
- Codeforces Elo
- A competitive programming rating derived from performance on Codeforces problems, used here as a proxy for a model's code generation and algorithmic reasoning capability.
- MathVision
- A benchmark that evaluates a model's ability to solve mathematical problems presented in visual or image-based formats.
- VideoReasonBench
- A benchmark that tests a model's reasoning capabilities over video content, with human-level performance used as a reference point.
- Vibe Coding
- An evaluation category measuring a model's ability to generate well-designed applications or code in response to short-turn, user-facing instructions, representing an economic-value task.
- Science Discovery
- An evaluation dimension in the Seed2.0 framework that probes a model's ability to perform research-style reasoning and code synthesis for scientific problems.
- Context Learning
- An evaluation dimension that assesses a model's ability to learn from and reason over information provided within a long input context.
- VQA (Visual Question Answering)
- A task in which a model must answer natural-language questions about the content of an image or visual input.
- MAE (Mean Absolute Error)
- A metric measuring the average absolute difference between predicted and actual values, used here to quantify counting accuracy on the FSC-147 benchmark.
- STEM benchmarks
- Evaluation datasets covering Science, Technology, Engineering, and Mathematics tasks, used to assess a model's technical and scientific reasoning abilities.
- tiered deployment
- A strategy of offering multiple model sizes (Pro, Lite, Mini) so that users can select the appropriate balance of capability, latency, and cost for their specific application.