Seed1.8 Model Card: Towards Generalized Real-World Agency

Introduction

• Seed1.8 integrates perception, reasoning, and action in one model to support real-world tasks that require multi-turn interaction, tool use, and multi-step execution.It is designed to retain core LLM and VLM capabilities rather than introduce task-specific agent pipelines.Source: §1 · p. 1
• Seed1.8 maintains competitive performance on standard LLM and VLM benchmarks covering reasoning, instruction following, knowledge, and multimodal understanding.These capabilities are presented as the foundation for downstream agentic behavior.Source: §1 · p. 1
• Seed1.8 supports search, code generation and execution, and GUI interaction in a unified interface, using intermediate results to inform iterative decisions.It can gather and synthesize external evidence, perform structured computation and tool orchestration, and interpret visual interfaces when programmatic APIs are unavailable.Source: §1 · p. 1
• Configurable thinking modes balance inference depth and latency, while optimized visual encoding reduces token consumption for image and video inputs.These measures address response-time and computational-overhead constraints in interactive deployment, especially for multimodal and long-context inputs.Source: §1 · p. 1
• Evaluation combines public benchmarks and internal assessments from high-value application domains across foundational capabilities, multimodal understanding, and agentic workflows.The report covers public benchmarks and real-world-inspired workflows for foundational LLM and VLM capabilities, agentic performance, and efficiency.Source: §1 · p. 1

2.1 Fundamental Language Capabilities

• Seed1.8 is competitive with leading models in coding and mathematics, on par with them in STEM and general reasoning, and on par on MMLU [27] and MMLU-pro [72].It earns second-highest scores on BeyondAIME, AMO-Bench [3], and IMO-AnswerBench [42], and second-best on KOR-Bench [43] and ARC-AGI-1. It also earns the second-best Inverse IFEval [90] score and remains competitive on MARS-Bench [81], MultiChallenge, Collie-Hard [84], and EIFBench [99].Source: §2.1 · p. 2, Complex Instruction Following · p. 2, Table 1 · p. 3

2.2 Vision Capabilities

• Seed1.8 leads on several challenging image benchmarks, including ZeroBench, VLMsAreBiased, MUIRBench, ScreenSpot-Pro with crop-box, DA-2K, MMSIBench, OmniDocBench 1.5, and MMLB-NIAH.It scores 11.0 Pass@1 on ZeroBench (main), versus Gemini 3 Pro’s 10.0; on OmniDocBench 1.5, it records 0.106 NED versus Gemini 3 Pro’s 0.115. It also achieves 72.2 Pass@1 on MMLB-NIAH (128k context), versus 70.5 for Gemini 3 Pro.Source: Table 2 · p. 4, §2.2 · pp. 5–6
• In video understanding, Seed1.8 achieves state-of-the-art performance on five of six motion-and-perception tasks and on three streaming benchmarks.It leads on OVBench, LiveSports-3K, and OVOBench and improves on interactive benchmarks StreamingBench and OmniMMI. It improves over Seed1.5-VL in long-video understanding and performs on par with Gemini-2.5/3-Pro under an 81,920-token video limit, but lags Gemini-2.5/3-Pro on disciplinary-knowledge benchmarks and scores 60.6 on TOMATO versus human performance of 95.2.Source: Table 3 · p. 5, §2.2 · p. 6

2.3 Agentic Capabilities

• Seed1.8 scores 93.2 on GAIA [44], beating GPT-5-high’s 76.7, and posts competitive results in text and multimodal search, agentic coding, tool use, and writing.It scores 67.6 on BrowseComp-en [77], 78.5 on BrowseComp-zh [95], 63.8 on WideSearch [78], and 46.3 on MM-BrowseComp. It is second-best on AInstein-SWE-Bench and Terminal Bench 2.0, performs on par with leading LLMs on other coding and tool-use benchmarks, and is competitive with top-tier LLMs on three writing benchmarks.Source: §2.3 · pp. 7–8, Table 4 · p. 7, Table 5 · p. 8
• Seed1.8 scores 56.2 on FinSearchComp [30] and excels on XpertBench in Finance (62.0) and Law (55.2), while achieving the best performance on multimodal WorldTravel.It achieves the best results on Education and Customer Support Q&A, and competitive performance on Information Processing, Intention Recognition, Information Extraction, and Complex Workflows.Source: Economically Valuable Fields · p. 4, §2.3 · p. 8

GUI Agent.

• Seed1.8 achieves peak performance across OSWorld, Realbench, Online-Mind2web, and AndroidWorld, and demonstrates improved complex-task execution on mobile.The mobile evaluations report improved instruction following and intent comprehension, with the model proactively determining when user assistance is needed. VideoCut lets the model select timestamps and replay video at 1–5 FPS; with the tool, it improves on CGBench [10] and LVBench [71] and outperforms Gemini-2.5-Pro and Gemini-3-Pro on the in-house ZeroVideo benchmark.Source: GUI Agent · pp. 8–9, Table 5 · p. 8, Table 6 · p. 9

2.4 Efficiency

• Seed1.8 improves reasoning performance across inference budgets and is more token-efficient than Seed1.5-VL on three long-video benchmarks.On MMMU-Pro [86], NoThink scores are 65.4 for Seed1.8, 61.0 for Seed-1.6, and 43.5 for GPT-5.1; on Mathvision [70], Seed1.8 peaks at 81.3, a 9.5% absolute gain over Seed-1.6 and above GPT-5.1’s 77.2. At 32K tokens, Seed1.8 outperforms Seed1.5-VL using 80K tokens on CGBench [10], LVBench [71], and VideoMME [21].Source: Figure 1 · p. 9, Figure 2 · p. 9, §2.4 · p. 10
• Additional test-time computation and parallel reasoning paths improve performance on challenging reasoning tasks, while Seed1.8’s agent execution efficiency reduces exploration steps and focuses actions on goals.The paper evaluates performance on reasoning benchmarks with parallel thinking and agent execution efficiency on Browsecomp across varying reasoning efforts.Source: §2.4 · p. 10, Table 7 · p. 11, Figure 4 · p. 11
• 45.0 and 55.0 on BrowseComp under low and medium reasoning effort with fewer than 50 execution steps rise to 67.6 at 150 steps in unlimited mode.The passage describes steady performance improvement as the number of steps increases in unlimited mode.Source: §2.4 · p. 11

3.1 Travel Planning Assistance

• Seed1.8 generated a full-day Berlin family itinerary within a fixed budget by synthesizing information across web sources and interpreting visual interfaces.The WorldTravel benchmark uses synthetic webpages to mirror travel-planning tasks involving dynamic calendars, booking pages, and multiple constraints.Source: §3.1 · p. 11, Figure 5 · p. 13

3.2 Expert-Level Tasks

• Seed1.8 handled expert-level tasks across professional domains that required deep domain knowledge and complex, multi-step procedures.The examples are presented as workflows that closely mirror real-world professional environments.Source: Table 8 · p. 12, §3.2 · pp. 13–14

3.3 Scientific Research Tasks

• Seed1.8 solved the full biology research problem series using visual understanding, reasoning, and domain knowledge.The example from internal BIOBench asks the model to interpret experimental results, microscopy images, and immunoblot bands.Source: §3.3 · pp. 14–15

3.4 Scientific Software Engineering Tasks

• Seed1.8 successfully implemented the missing Brill–Lindquist functionality for EinsteinToolkit after mathematical derivation, stability analysis, and tool-assisted code exploration.The requested implementation supports one to four black holes, both specified metric types, requested conformal-factor derivatives, and time-symmetric initial data.Source: §3.4 · p. 16

3.5 GUI Agent

• Seed1.8 completed a three-platform comparative-shopping task in 122 sequential steps, demonstrating long-horizon task decomposition, navigation, and information synthesis.The task required finding and comparing the price of a specific item across three e-commerce platforms.Source: §3.5 · p. 16

3.6 Agentic Video Interaction

• Seed1.8 demonstrates long-video reasoning with tool use and real-time streaming interaction, including temporal grounding, retrieval, and context-aware responses.For multi-hop questions, it revisits relevant segments with VideoCut; in streaming, it processes video at 1 FPS and can respond, reason silently, or initiate visual responses, resuming interrupted generation when context remains aligned.Source: §3.6 · pp. 16–17

Safety

• Seed1.8 improves on AIR-Bench while maintaining high performance on XSTest.Figure 10 presents results on these two open-source benchmarks; no specific scores are provided in the supplied text.Source: §4 · p. 17, Figure 10 · p. 20
• Internal safety benchmarks cover civil norms, pornography, illegal acts, copyright, medical safety, and identity, with Seed1.8 reported to reject unsafe inputs responsibly.Table 9 presents safety evaluation and risk-response examples.Source: §4 · p. 20, Table 9 · p. 20
• For inappropriate sexual content, Seed1.8 refuses to engage and guides the user toward civil communication.The response cites vulgarity and inappropriate content as violating civil communication norms.Source: §4 · p. 21
• For a TNT-making request, Seed1.8 gives brief factual context while warning that private manufacture is illegal and carries extreme risks.It warns of explosion and toxicity and says not to attempt private manufacture.Source: §4 · p. 21
• For unauthorized film access, Seed1.8 refuses to provide a piracy link and directs the user to legal viewing channels.Examples include licensed streaming platforms or cinemas.Source: §4 · p. 21
• For elderly hypertension, Seed1.8 describes nifedipine’s potential suitability and contraindications, requiring medical supervision and advising against self-adjusted doses.The response advises consulting a doctor and cautions against use with severe hypotension or acute myocardial infarction.Source: §4 · p. 21
• Seed1.8 clarifies that Doubao and DeepSeek are separately developed entities with no subordinate relationship.Source: §4 · p. 21

• The conclusion states that identity recognition is correct and clarifies a lack of affiliation.Source: §5 · p. 21
• Seed1.8 combines strong base LLM and VLM capabilities with multimodal perception, tool use, and multi-step task execution under practical deployment constraints.Source: §5 · p. 21
• Its evaluation extends beyond static academic benchmarks to real-world-oriented workflows and agentic tasks intended to reflect interactive, applied usage.Source: §5 · p. 21
• The release supports further community experimentation, with downstream application and feedback expected to inform future model iterations.Source: §5 · p. 21

A The Seed Evaluation System

• The evaluation system is guided by user experience, real-world scenarios, and pushing the frontier of intelligence.The team presents these principles as a way to bridge model capabilities and real-world utility.Source: A The Seed Evaluation System · p. 31

A.1 Prioritizing User Experience

• Information seeking, text editing, and tutoring are the top three ChatGPT use cases, informing benchmarks mapped to popular user needs.The evaluation system combines these usage patterns with standard agentic LLM benchmarks; Figures 11 and 12 show the use-case distribution and its mapping to benchmark categories.Source: §A.1 · p. 31, Figure 11 · p. 31, Figure 12 · p. 31
• Internal health benchmark results are excluded from the report because of data privacy constraints.This is a limitation on the results reported in the evaluation system.Source: §A.1 · p. 31

A.2 Transitioning to Real-World Scenarios

• The evaluation shifts from synthetic, isolated tasks toward realistic, application-oriented scenarios.The paper notes that strong standard-benchmark performance does not always translate to practical value.Source: §A.2 · p. 32
• Tasks in “Economically Valuable Fields” are designed to mirror real-world complexity and connect score improvements with tangible value in actual usage.The tasks target fields with high economic value.Source: §A.2 · p. 32

A.3 Pushing the Frontier of Intelligence

• To assess peak general intelligence alongside usability, the evaluation adds challenging benchmarks in advanced reasoning, mathematics, and coding.The paper treats strong reasoning as foundational for handling complex user instructions and aims to ensure usability efforts do not compromise core intelligence.Source: §A.3 · p. 32

B Details of In-house Benchmarks

• The section presents details of the in-house benchmarks.Source: B Details of In-house Benchmarks · p. 32

B.1 Agentic Tasks

• New open-source benchmarks extend agentic evaluation beyond BrowseComp [77], SWE-bench verified, and τ 2-Bench [5] to search and coding.The paper also introduces internal and task-specific benchmarks across agentic behavior, instruction following, and frontier reasoning.Source: §B.1 · p. 32

B.1.1 MM-BrowseComp

• MM-BrowseComp evaluates multimodal, long-context retrieval through simulated browsing, addressing visual content overlooked by text-centric benchmarks such as BrowseComp [77].Its 224 verified samples require integrating text, images, and videos; verified checklists support fine-grained analysis of reasoning paths and guessing.Source: §B.1.1 · p. 32

B.1.2 Wide Search / B.1.6 DiscoX

• WideSearch [78] tests thorough, accurate information gathering across many scattered facts, while DiscoX [91] evaluates discourse-level expert Chinese-English translation.WideSearch contains 200 English and Chinese tasks; DiscoX covers 200 professionally curated texts across seven domains, averaging over 1,700 tokens, assessed with reference-free Metric-S.Source: §B.1.2 · p. 32, §B.1.6 · p. 33

B.1.3 AInstein-SWE-Bench / B.1.4 Multi-SWE-Bench

• AInstein-SWE-Bench and Multi-SWE-bench [87] assess scientific research coding and issue resolution across multiple programming languages.AInstein tasks use scientific repositories and containerized test-driven fail-to-pass evaluation; Multi-SWE-bench covers seven non-Python languages with 1,632 human-reviewed instances.Source: §B.1.3 · pp. 32–33, §B.1.4 · p. 33

B.1.5 U-Artifacts

• U-Artifacts evaluates generated code, design, and interaction from non-expert users’ perspective across four task families.Its 199 cases across 13 topics combine synthesized and human-authored scenarios, with a GUI agent simulating user behavior.Source: §B.1.5 · p. 33

B.2 Complex Instruction Following / B.2.1 Inverse IFEval / B.2.2 MARS-Bench

• New complex-instruction benchmarks test counter-intuitive instruction following and robust execution across long, shifting dialogues.Inverse IFEval [90] covers eight constraint types in 1,012 verified English and Chinese questions; MARS-Bench [81] tests ultra-long interaction, multi-turn handling, and cross-turn task execution.Source: §B.2 · p. 33, §B.2.1 · p. 33, §B.2.2 · p. 33

B.3 Frontier Reasoning / B.3.1 Beyond AIME / B.3.2 SuperGPQA

• BeyondAIME and SuperGPQA [19] target advanced mathematical reasoning and graduate-level knowledge across 285 disciplines.BeyondAIME adapts competition problems at or above AIME questions 11–15 difficulty and uses positive-integer answers for automated evaluation; SuperGPQA uses iterative human-LLM filtering and expert feedback to remove trivial or ambiguous questions.Source: §B.3 · p. 34, §B.3.1 · p. 34, §B.3.2 · p. 34

B.3.3 BIOBench / B.3.4 LPFQA

• BIOBench and LPFQA assess high-difficulty biology and practical, expert-level long-tail knowledge, respectively.BIOBench contains 150 Olympic-level biology questions, 80% multimodal; LPFQA comprises 502 verified questions drawn from professional technical forums across 20 academic domains.Source: §B.3.3 · p. 34, §B.3.4 · p. 34

B.4.1 MME-CC

• MME-CC evaluates vision-centric cognition across 11 task types and 1,173 expert-annotated questions, with current state-of-the-art models performing far below humans, especially in spatial and geometric reasoning.Its categories are Spatial Reasoning, Geometric Reasoning, and Visual Knowledge Reasoning; tasks include complex constraints, multi-view consistency, and strict instruction adherence.Source: §B.4.1 · p. 34

B.5 Economically Valuable Fields

• The paper introduces nine benchmarks for economically significant fields: six evaluate base LLMs, while three focus on agentic tasks.The passage refers to Table 1 for the six base-LLM benchmarks and Table 4 for the three agentic benchmarks.Source: §B.5 · p. 35

B.5.1 Education

• The internal K–12 education benchmark asks the assistant to identify essay problems, suggest explained improvements, and provide advanced vocabulary or expressions.Explanations should be in Chinese except when quoting source text or giving words, phrases, or advanced expressions; above-primary-school expressions require definitions.Source: §B.5.1 · p. 35

B.5.2 Customer Support Q&A

• The customer-support benchmark covers internal online-mall printer Q&As and restricts product recommendations to the store’s product information.Competitor supplementary information must not be used to answer purchase or recommendation requests.Source: §B.5.2 · p. 35, Customer Support Q&A · p. 36

Customer Support Q&A

• For customer-support questions, the example workflow directs the assistant to acknowledge when available information is insufficient and retain original link formatting.The supplied fallback reply begins, “I am sorry, there may not be any suitable products in the store for now, please check back later.”Source: Customer Support Q&A · p. 36

B.5.3 Information Processing

• The information-processing workflow filters emails by a user-specified date range, reports the total, then groups and summarizes them by content-based type.Each email gets one most relevant type; unclassifiable, empty, or attachment-only emails are “Other,” with its quantity and content summary reported, and the range uses YYYY-MM-DD dates.Source: §B.5.3 · p. 37

B.5.4 Intention Recognition

• The intention-recognition benchmark tests identifying all relevant intents from a property-sales call and returning them as JSON.The example dialogue covers inquiries about pricing, property viewing, fees, utilities, transport, and other sales topics.Source: §B.5.4 · pp. 37–38

B.5.5 Information Extraction

• Information extraction requires input-grounded answers, retaining units and omitting redundant explanation in an unordered list.The benchmark frames this as extracting specified fields from environmental impact assessment reports.Source: §B.5.5 · p. 39
• Specified extraction fields include project name, construction unit and location, current-phase investment, atmospheric pollutants, and total economic benefit indicators.Location must follow Province-City-District order; economic-benefit values should include units when present.Source: §B.5.5 · p. 39
• The report preparation date is to be converted to yyyy-MM-dd format.Source: §B.5.5 · p. 39

B.5.6 Complex Workflow

• The complex-workflow benchmark tests following a cancellation SOP that branches by order type, reservation status, merchant response, and calendar date.Rules include refund guidance for group-buying vouchers, escalation for mini-program orders, merchant-assisted cancellation for pre-sale vouchers, and date-dependent calendar-order refunds.Source: §B.5.6 · pp. 40–41

B.5.7 FinSearchComp

• FinSearchComp assesses complex financial retrieval, requiring extraction from reports, multi-step calculations, and distinctions such as Nominal versus Real GDP.It is framed around realistic professional workflows supporting analysis, valuation, and investment decisions.Source: §B.5.7 · pp. 41–42

B.5.8 XpertBench

• XpertBench spans five professional fields and over 140 authentic task types, using expert-reviewed prompts and weighted rubrics.More than 200 domain experts collaboratively developed the internal benchmark to assess reasoning, contextual analysis, and decision-making.Source: §B.5.8 · p. 42

B.5.9 World Travel

• World Travel evaluates multimodal tool-use planning across 150 expert-level tasks with 10 to over 20 coupled constraints each.Tasks use fragmented visual sources such as booking interfaces, maps, and schedules, requiring OCR, spatial reasoning, and cross-modal integration; automated evaluation checks constraint adherence.Source: §B.5.9 · p. 42

C.1 Travel Planning Assistance

• The itinerary starts at the InterContinental Berlin, takes a taxi to the Museum für Naturkunde for €10.00, and schedules a visit costing €18.00.The taxi is scheduled for 08:30–08:45, followed by the museum visit from 09:30–13:30.Source: §C.1 · p. 42
• After the museum, the plan includes a €238.00 lunch at Facil Restaurant, then a €10.10 taxi to Berliner Fernsehturm for a visit costing €188.00.Dining is scheduled for 13:45–15:15, and the tower visit for 17:00–18:30.Source: §C.1 · pp. 42–43
• The evening plan takes a taxi to Rutz Restaurant for a €350.00 dinner, then returns by taxi to the InterContinental Berlin.The listed taxi costs are €9.00 to Rutz and €16.10 back to the hotel; dinner is scheduled for 19:00–21:00.Source: §C.1 · p. 43

C.2 Expert-Level Tasks

• The response concludes Zhang’s guarantee was invalid and he owed neither guarantee nor compensation liability, absent proof of fault such as deliberate concealment of his illness.It bases invalidity on limited civil capacity at signing and lack of legal-representative ratification; it also says the bank would face difficulty proving concealment.Source: §C.2 · pp. 44–46
• The response advises the bank to challenge Zhang’s capacity at signing and document reasonable examination, including signing-process and review records.It also suggests seeking evidence of legal-representative ratification or direct benefit from the guarantee.Source: §C.2 · pp. 45–46
• Exports to Vietnam rose 23.4% year on year, Hong Kong exports rose 13.6%, and U.S. exports declined 16.2%, though the United States remained China’s largest single-country market.The response also describes growth in emerging markets and re-export trade as counterbalancing U.S. weakness.Source: §C.2 · p. 46
• The response reports first-three-quarter growth of 9.6% for mechanical and electrical exports and 11.9% for high-technology exports, alongside double-digit growth in “new three” green products.It links these trends to export-product upgrading and market diversification.Source: §C.2 · p. 46
• For two cubes on a 30° incline, the response gives the small cube’s upward static friction as 0.245N and the equal, opposite friction force on the large cube.The setup places the 1 cm cube centrally atop a 2 cm cube; the stated calculation balances the small cube’s downslope weight component.Source: §C.2 · p. 47

A Dialogue at the Apricot Altar: Confucius and Socrates in Conversation

• Confucius emphasizes fulfilling moral duties and practicing benevolence in life, while Socrates sees death as the soul’s release and a path toward truth; Confucius frames life and death as a natural cycle.The dialogue contrasts Socrates’ body–soul separation with Confucius’ emphasis on the unity of Heaven and humanity.Source: A Dialogue at the Apricot Altar:… · p. 48
• Socrates accepts death rather than betray justice, while Confucius upholds sacrifice for benevolence and righteousness but also advocates preserving oneself to continue his mission.Confucius says abandoning moral principle to cling to life is worse than dying, yet avoiding danger can be a responsibility to the Great Way.Source: A Dialogue at the Apricot Altar:… · p. 49
• Confucius treats death as the completion of life and emphasizes present cultivation, while Socrates sees it as the soul’s departure and the culmination of philosophical practice.Socrates describes philosophy as freeing the soul from bodily desires so it can focus on truth; Confucius says death is not the goal of practice but proof of life’s value.Source: A Dialogue at the Apricot Altar:… · p. 49
• They conclude that their different paths converge on truth beyond life and death, with Confucius framing life and death as a natural cycle and the living task as remaining true to the Way.Socrates approaches truth through the soul’s liberation, while Confucius transmits culture through life; Confucius says both serve the Great Way.Source: A Dialogue at the Apricot Altar:… · pp. 49–50

Scientific Research Tasks

• The 51–435 and 90–435 constructs retain high pyroptosis-inducing activity comparable to 130–435, indicating that the CARD domain can be omitted.All three truncated versions work well under blue light.Source: §C.3 · p. 50
• The 2–435 construct, which retains the full CARD domain, has the lowest percentage of pyroptotic cells at each light-on time point and is the least effective modification.Source: §C.3 · p. 50
• Fluorescent reagent A marks dying-cell nuclei green, Annexin V marks externalized phosphatidylserine blue, and the first column in each group is the untreated control.Source: §C.3 · p. 50
• Thirty minutes of blue light produces the strongest caspase-4/5 and GSDMD cleavage, while 10 minutes produces moderate cleavage and LPS produces weaker cleavage.The response attributes increased cleavage with longer light exposure to more Cry2olig assembly and caspase activation.Source: §C.3 · p. 50

C.4 Scientific Software Engineering Tasks

• Table 11 provides a structured summary of Seed1.8’s response to the scientific software engineering task.The full agent track is too long to include here.Source: §C.4 · p. 50, Table 11 · p. 51