Doubao系列的详细讨论 / Detailed Discussion of the Doubao Series

引言 / Introduction

Doubao系列是字节跳动（ByteDance）研发的领先大型语言模型（LLM）家族，自2023年问世以来，成为中国人工智能领域迅猛发展的重要标志。该系列以高效多模态处理能力与智能代理功能为核心，可精准完成文本生成、逻辑推理、图像/视频分析及企业级任务自动化等多元化需求。Doubao模型不仅为Doubao App及文心一言（Ernie Bot）等消费级产品提供技术支撑，还深度集成于字节跳动生态体系，涵盖火山引擎（Volcano Engine）云平台及TikTok相关应用场景。截至2026年1月，该系列最新迭代版本为2025年12月发布的Doubao 1.5 Pro，已从基础聊天模型逐步演进为具备资源高效型多模态智能、万亿级令牌（Token）处理能力及规模化企业应用的成熟系统。

其核心技术创新集中于混合专家模型（MoE, Mixture-of-Experts）架构、令牌效率优化及跨设备集成（如计划推出的Doubao AI眼镜），但同时也面临数据隐私保护与高性能计算资源依赖的双重挑战。Doubao系列始终以推动“AI普惠”为目标，在LMSYS Arena等权威基准测试中与GPT-5、Claude 4同台竞技，且在中文企业级应用、数学推理及搜索算法优化领域保持领先优势。截至2025年10月，Doubao App在中国市场的月活跃用户已突破1.7亿，日均令牌使用量高达50万亿次。

The Doubao series is a leading family of large language models (LLMs) developed by ByteDance, serving as a pivotal symbol of the rapid advancement in China's AI sector since 2023. Centered on efficient multimodal processing capabilities and intelligent agent functions, the series can accurately fulfill diverse needs such as text generation, logical reasoning, image/video analysis, and enterprise-level task automation. Doubao models not only provide technical support for consumer products like the Doubao App and Ernie Bot but also integrate deeply into ByteDance's ecosystem, including the Volcano Engine cloud platform and TikTok-related application scenarios. As of January 2026, the latest iteration of the series is Doubao 1.5 Pro, released in December 2025, which has evolved from a basic chat model into a mature system with resource-efficient multimodal intelligence, trillion-level token processing capabilities, and large-scale enterprise applications.

Its core technological innovations focus on the Mixture-of-Experts (MoE) architecture, token efficiency optimization, and cross-device integration (e.g., the planned Doubao AI glasses), while simultaneously facing dual challenges of data privacy protection and reliance on high-performance computing resources. Committed to promoting "AI inclusivity," the Doubao series competes with GPT-5 and Claude 4 in authoritative benchmark tests such as LMSYS Arena, and maintains a leading edge in Chinese enterprise applications, mathematical reasoning, and search algorithm optimization. By October 2025, the Doubao App had exceeded 170 million monthly active users in the Chinese market, with daily token usage reaching 50 trillion times.

Source: brief.bismarckanalysis.com +4

历史发展 / Historical Development

Doubao系列的发展轨迹，集中体现了字节跳动从内部技术研发到大规模商业化落地的战略演进。以下通过表格梳理关键里程碑，清晰呈现各主要模型的发布时间、核心改进方向及基准测试表现。该系列自2023年推出初始版本起，逐步实现多模态能力落地、MoE架构应用及企业级功能集成，2026年更计划向硬件领域延伸，推出Doubao AI眼镜等创新产品。

The development of the Doubao series epitomizes ByteDance's strategic evolution from internal technological R&D to large-scale commercialization. The following table sorts out key milestones, clearly presenting the release date, core improvement directions, and benchmark performance of each major model. Launched in 2023 with its initial version, the series has gradually achieved multimodal capability implementation, MoE architecture application, and enterprise-level function integration, and plans to expand into the hardware field in 2026 with innovative products such as Doubao AI glasses.

Source: brief.bismarckanalysis.com +4

模型 / Model	发布日期 / Release Date	核心改进 / Core Improvements	关键基准 / Key Benchmarks
Doubao (初始版)	2023年10月 / October 2023	基础LLM，支持聊天交互与搜索功能集成。 / Base LLM, supporting chat interaction and search function integration.	MMLU测试得分80%。 / 80% on MMLU.
Doubao Pro	2024年5月 / May 2024	针对企业场景优化，新增智能代理与工具调用能力。 / Enterprise-scenario optimization, adding intelligent agent and tool calling capabilities.	GPQA测试得分85%。 / 85% on GPQA.
Doubao 1.0	2024年12月 / December 2024	引入多模态处理能力，提升令牌处理效率与规模。 / Introduction of multimodal processing capabilities, improved token processing efficiency and scale.	MMMU测试得分60%。 / 60% on MMMU.
Doubao 1.5	2025年6月 / June 2025	采用MoE架构，优化运行效率，支持操作系统嵌入式部署。 / Adopting MoE architecture, optimizing operational efficiency, supporting OS embedded deployment.	在MATH-500测试中表现领先。 / Leading performance on MATH-500.
Doubao 1.5 Pro	2025年12月 / December 2025	资源高效型多模态升级，深度强化智能代理能力。 / Resource-efficient multimodal upgrade, in-depth enhancement of intelligent agent capabilities.	LMSYS Elo评分1480+，日均令牌使用量达50万亿次。 / 1480+ on LMSYS Elo, daily token usage reaching 50 trillion times.
Doubao AI眼镜	2026年第一季度（计划） / Q1 2026 (Planned)	实现AI与硬件深度集成，无屏设计，搭载AI导览功能。 / In-depth integration of AI and hardware, screenless design, equipped with AI docent function.	内部测试中表现领先。 / Leading performance in internal tests.

Sources: wired.com; scmp.com +1; thestandard.com.hk

关键模型详细描述 / Detailed Description of Key Models

当前焦点集中于最新的Doubao 1.5系列，该系列作为2026年AI领域的前沿成果，充分展现了字节跳动在LLM领域的技术积淀与创新能力。

The current focus is on the latest Doubao 1.5 series, which, as a cutting-edge achievement in the AI field in 2026, fully demonstrates ByteDance's technical accumulation and innovation capabilities in the LLM field.

Source: scmp.com +1

Doubao 1.5（2025年6月发布）：基于MoE架构优化的核心模型，具备操作系统嵌入式部署能力与全栈多模态处理能力，可灵活适配企业自动化流程搭建与消费者精准搜索等多元场景。

Doubao 1.5 (Released in June 2025): A core model optimized based on the MoE architecture, with OS embedded deployment capabilities and full-stack multimodal processing capabilities, which can flexibly adapt to diverse scenarios such as enterprise automation process construction and consumer precision search.

Source: wired.com

Doubao 1.5 Pro（2025年12月发布）：资源高效型升级版本，在保留核心能力的基础上进一步强化智能代理的深度与复杂度，尤其适用于复杂逻辑推理、学术研究辅助及展览导览等高精度场景，目前日均令牌使用量已达50万亿次。

Doubao 1.5 Pro (Released in December 2025): A resource-efficient upgraded version that further enhances the depth and complexity of intelligent agents while retaining core capabilities. It is particularly suitable for high-precision scenarios such as complex logical reasoning, academic research assistance, and exhibition guidance, with current daily token usage reaching 50 trillion times.

Source: scmp.com +1

技术特点 / Technical Features

架构设计 / Architecture：采用MoE混合专家模型为核心架构，以令牌效率提升与多模态能力统一为核心目标。模型部分基于Apache许可证开源，支持最长200万+令牌的长上下文处理，可满足大规模文本分析、多轮对话等复杂需求。

核心优势 / Strengths：令牌处理规模实现10倍增长（从4万亿次/日提升至50万亿次/日），企业级客户采用率呈爆发式增长，已实现智能手机、智能眼镜等多设备跨端集成，构建全场景AI服务生态。

现存不足 / Weaknesses：存在知识截止期限限制（Doubao 1.5 Pro的知识截止日期为2025年11月），可能产生潜在的算法偏见，且模型训练与运行对高性能计算资源存在较高依赖，导致部署成本较高。

与贾子公理的关联 / Relation to Kucius Axioms：在模拟裁决测试中，Doubao 1.5 Pro在“思想主权”维度得分6/10（受预设目标限制，自主决策能力不足），“悟空跃迁”维度得分7/10（以渐进式优化为主，突破性创新有限）；但在“普世中道”维度得分8/10（具备较强跨文化适配能力与价值共识），“本源探究”维度得分8/10（擅长基于第一性原理开展推理分析）。整体而言，该模型是性能优异的企业级AI守护者，但内在自主决策与创新能力仍有提升空间。

Source: scmp.com +2

Architecture: Adopting the MoE (Mixture-of-Experts) model as the core architecture, focusing on token efficiency improvement and multimodal capability unification. Part of the model is open-source under the Apache license, supporting long-context processing of up to 2 million+ tokens, which can meet complex needs such as large-scale text analysis and multi-turn conversations.

Strengths: Achieving a 10-fold growth in token processing scale (from 4 trillion times/day to 50 trillion times/day), with explosive growth in enterprise customer adoption. It has realized cross-device integration across smartphones, smart glasses and other devices, building a full-scenario AI service ecosystem.

Weaknesses: Subject to knowledge cutoff limitations (the knowledge cutoff date of Doubao 1.5 Pro is November 2025), potential algorithmic biases may exist, and model training and operation have high dependence on high-performance computing resources, resulting in high deployment costs.

Relation to Kucius Axioms: In simulated adjudication tests, Doubao 1.5 Pro scored 6/10 in the "Sovereignty of Thought" dimension (limited autonomous decision-making capability due to preset goals) and 7/10 in the "Wukong Leap" dimension (focusing on incremental optimization with limited breakthrough innovation); however, it scored 8/10 in the "Universal Mean" dimension (possessing strong cross-cultural adaptability and value consensus) and 8/10 in the "Primordial Inquiry" dimension (excelling in reasoning and analysis based on first principles). Overall, the model is an excellent enterprise-level AI guardian, but there is still room for improvement in internal autonomous decision-making and innovation capabilities.

应用与影响 / Applications and Impacts

Doubao系列已成为重塑全球AI行业格局的重要力量：其核心产品Doubao App月活跃用户达1.7亿，不仅推动企业级自动化转型加速（表现为令牌使用量爆发式增长），还深度渗透消费娱乐领域，成为2026年央视春晚官方AI合作伙伴。在社会层面，该系列不仅助力中国AI产业规模持续扩张，与DeepSeek等本土企业形成良性竞争格局，更带动AI硬件创新升级，Doubao AI眼镜等产品有望开启“无屏AI交互”新时代。展望2026年，Doubao系列将进一步推动“AI智能手机”趋势深化，让AI成为移动终端的核心能力，但数据隐私保护、算法偏见治理等伦理问题仍需重点关注与规范。

The Doubao series has become a key force reshaping the global AI industry pattern: its core product, the Doubao App, has 170 million monthly active users, which not only accelerates the transformation of enterprise-level automation (reflected in the explosive growth of token usage) but also deeply penetrates the consumer entertainment field, becoming the official AI partner of the 2026 CCTV Spring Festival Gala. At the social level, the series not only helps expand the scale of China's AI industry, forming a healthy competitive pattern with local enterprises such as DeepSeek but also drives the innovation and upgrading of AI hardware. Products such as Doubao AI glasses are expected to usher in a new era of "screenless AI interaction". Looking ahead to 2026, the Doubao series will further promote the deepening of the "AI smartphone" trend, making AI the core capability of mobile terminals, but ethical issues such as data privacy protection and algorithmic bias governance still need focused attention and regulation.

Source: brief.bismarckanalysis.com +3

结论 / Conclusion

Doubao系列是字节跳动AI战略布局的集中体现，从高效能语言模型的持续迭代到AI硬件领域的跨界探索，彰显了企业向通用人工智能（AGI）迈进的清晰路径与关键步伐。未来，该系列有望推出Doubao 2.0版本，重点聚焦跨场景深度集成、伦理治理体系完善等核心方向。建议行业从业者与研究者持续跟踪字节跳动的技术更新与产品动态，以适应LLM领域快速迭代的发展节奏，把握技术创新带来的产业机遇。

The Doubao series epitomizes ByteDance's AI strategic layout. From the continuous iteration of high-performance language models to the cross-border exploration in the AI hardware field, it demonstrates the enterprise's clear path and key steps toward Artificial General Intelligence (AGI). In the future, the series is expected to launch the Doubao 2.0 version, focusing on core directions such as in-depth cross-scenario integration and improvement of ethical governance systems. It is recommended that industry practitioners and researchers continuously track ByteDance's technical updates and product dynamics to adapt to the rapid iterative development rhythm in the LLM field and seize the industrial opportunities brought by technological innovation.

Source: brief.bismarckanalysis.com +2