核心内容摘要
三生有幸，此“家”有你：一场温暖生日，道尽三姐妹情深
ERNIE系列的详细讨论 / Detailed Discussion of the ERNIE Series引言 / IntroductionERNIEEnhanced Representation through kNowledge IntEgration系列是由百度开发的知识增强预训练语言模型LLM家族自2019年问世以来成为中国人工智能领域取得重大进步的标志性成果。
该系列模型以知识图谱集成与连续学习为核心技术支撑具备处理文本、图像、视频等多类型数据及多模态任务的综合能力。
ERNIE模型不仅是Ernie Bot文心一言聊天机器人的核心驱动力还深度融入百度全生态体系广泛应用于千帆AI云平台Qianfan AI Cloud Platform、百度搜索等核心服务中。
截至2026年1月该系列最新迭代版本为2025年12月发布的ERNIE
0已从最初的基础自然语言处理NLP模型演进为具备高级推理、多模态合成能力及企业级规模化应用价值的综合性AI系统。
其核心创新点集中在三大维度一是知识增强机制的持续迭代实现知识与数据的深度融合二是参数规模的突破性扩展峰值达
4万亿大幅提升模型性能上限三是开放开源策略的落地推动生态协同发展。
同时该系列模型也面临数据隐私保护、超高计算成本两大核心挑战。
ERNIE系列的核心愿景是推动“知识驱动型AI”的发展在LMSYS Arena等国际权威基准测试中与GPT、Gemini等顶尖模型展开激烈竞争尤其在中文NLP任务、图像分析、视频理解三大领域保持领先优势。
The ERNIE (Enhanced Representation through kNowledge IntEgration) series is a family of knowledge-enhanced pre-trained language models (LLMs) developed by Baidu, marking significant advancements in Chinas AI landscape since
Centered on knowledge graph integration and continual learning, the series is capable of processing text, images, videos, and undertaking multimodal tasks. ERNIE models not only power the Ernie Bot (Wenxin Yiyan) chatbot but also are widely integrated into Baidus ecosystem, including the Qianfan AI cloud platform and search services. As of January 2026, the latest model in the series is ERNIE
0 (released in December
, which has evolved from a basic NLP model to a comprehensive system with advanced reasoning capabilities, multimodal synthesis functions, and enterprise-level application potential.Its core innovations lie in three key aspects: first, the continuous iteration of knowledge enhancement mechanisms to achieve in-depth integration of knowledge and data; second, the breakthrough expansion of parameter scale, peaking at
4 trillion, which greatly raises the upper limit of model performance; third, the implementation of open-source strategies to promote ecological collaborative development. Meanwhile, the series also faces two core challenges: data privacy protection and ultra-high computing costs. The core vision of the ERNIE series is to advance the development of knowledge-driven AI. In international authoritative benchmarks such as LMSYS Arena, it competes fiercely with top models like GPT and Gemini, maintaining leading advantages particularly in three fields: Chinese NLP tasks, image analysis, and video understanding.历史发展 / Historical DevelopmentERNIE系列的发展历程清晰映射了人工智能从知识增强型自然语言处理向多模态智能演进的核心趋势。
以下通过表格梳理各关键里程碑详细呈现主要模型的发布时间、核心改进方向及关键基准测试表现。
该系列自ERNIE
0开启知识增强探索之路逐步实现多模态能力落地、开源生态构建及参数规模跨越式增长截至2026年ERNIE
0已成为全球多模态AI领域的前沿代表。
The development of the ERNIE series clearly reflects the core trend of artificial intelligence evolving from knowledge-enhanced natural language processing to multimodal intelligence. The following table sorts out key milestones, detailing the release date, core improvement directions, and key benchmark performance of each major model. Starting with ERNIE
0s exploration of knowledge enhancement, the series has gradually achieved the implementation of multimodal capabilities, the construction of an open-source ecosystem, and the leapfrog growth of parameter scale. As of 2026, ERNIE
0 has become a frontier representative in the global multimodal AI field.模型 / Model发布日期 / Release Date核心改进 / Core Improvements关键基准 / Key BenchmarksERNIE
02019年3月 / March 2019首次集成知识图谱实现实体与关系的精准建模。
/ First integration of knowledge graphs to achieve accurate modeling of entities and relationships.在中文GLUE任务中表现领先。
/ Leading performance in Chinese GLUE tasks.ERNIE
02020年7月 / July 2020提出连续学习框架构建多任务预训练体系。
/ Proposed a continual learning framework and built a multi-task pre-training system.在SuperGLUE基准测试中实现性能显著提升。
/ Significant performance improvement in SuperGLUE benchmarks.ERNIE
02021年7月 / July 2021参数规模扩展至100亿新增知识增强生成能力。
/ Expanded parameter scale to 10 billion and added knowledge-enhanced generation capabilities.在FewGLUE任务中达成90%以上准确率。
/ Achieved over 90% accuracy in FewGLUE tasks.ERNIE
52023年6月 / June 2023支持插件扩展功能成为Ernie Bot的核心基础模型。
/ Supported plug-in expansion and became the core base model for Ernie Bot.在MMLU基准测试中获得85%准确率。
/ Achieved 85% accuracy in MMLU benchmarks.ERNIE
02023年10月 / October 2023突破单模态局限实现多模态理解与图像生成能力。
/ Broke the limitation of unimodality and achieved multimodal understanding and image generation capabilities.在GPQA基准测试中达到80%准确率。
/ Achieved 80% accuracy in GPQA benchmarks.ERNIE
0 Turbo2024年6月 / June 2024聚焦速度优化实现实时响应能力适配低延迟场景。
/ Focused on speed optimization, achieved real-time response capabilities, and adapted to low-latency scenarios.LMSYS Elo评分突破1300分。
/ LMSYS Elo score exceeded
ERNIE
52025年6月 / June 2025推出开源混合专家MoE模型家族含10个变体实现多模态统一建模。
/ Launched an open-source Mixture of Experts (MoE) model family with 10 variants, enabling unified multimodal modeling.在MMMU基准测试中获得62%准确率。
/ Achieved 62% accuracy in MMMU benchmarks.ERNIE
02025年12月 / December 2025参数规模达
4万亿采用原生全模态建模技术强化图像/视频分析能力。
/ Parameter scale reached
4 trillion, adopted native full-modality modeling technology, and enhanced image/video analysis capabilities.AIME基准测试准确率超95%SWE-Bench基准测试准确率达80%。
/ Over 95% accuracy in AIME benchmarks and 80% accuracy in SWE-Bench benchmarks.ERNIE Next Gen2025年3月中旬计划 / Mid-March 2025 (Planned)重点提升推理能力与领域适配能力优化垂直场景表现。
/ Focus on improving reasoning capabilities and domain adaptation, optimizing performance in vertical scenarios.内部测试中表现领先同类型模型。
/ Leading performance among similar models in internal tests.从ERNIE
0的实验性探索到ERNIE
0的商业化成熟应用该系列模型的参数规模从数十亿量级跃升至万亿量级深刻标志着人工智能从“知识增强”向“全模态智能”的战略转型。
截至2026年1月Ernie Bot的月活跃用户MAU已突破2亿成为全球用户规模领先的AI对话产品之一。
From the experimental exploration of ERNIE
0 to the mature commercial application of ERNIE
0, the parameter scale of the series has jumped from billions to trillions, profoundly marking the strategic transformation of artificial intelligence from knowledge enhancement to full-modality intelligence. As of January 2026, Ernie Bots Monthly Active Users (MAU) have exceeded 200 million, making it one of the AI dialogue products with the leading user scale in the world.关键模型详细描述 / Detailed Description of Key Models作为2026年ERNIE系列的前沿代表ERNIE
5与ERNIE
0凭借突破性技术与规模化应用能力成为行业关注的核心焦点二者共同构建了百度多模态AI的技术底座。
As the frontier representatives of the ERNIE series in 2026, ERNIE
5 and ERNIE
0 have become the core focus of the industry with their breakthrough technologies and large-scale application capabilities, jointly building the technical foundation of Baidus multimodal AI.ERNIE
52025年6月该模型以开源混合专家MoE模型家族为核心特色包含10个不同功能定位的变体模型可根据场景需求灵活适配。
其核心突破在于实现多模态统一建模打破文本、图像、音频等不同模态数据的处理壁垒实现跨模态信息的深度融合。
该模型已深度集成至Ernie Bot为其提供图像生成、复杂逻辑推理、长文本处理等核心能力同时依托开源策略赋能开发者生态推动中小企业及科研机构的AI创新应用。
ERNIE
5 (June
: Featuring an open-source Mixture of Experts (MoE) model family with 10 variants of different functional orientations, it can be flexibly adapted according to scenario requirements. Its core breakthrough lies in realizing unified multimodal modeling, breaking the processing barriers of different modal data such as text, images, and audio, and achieving in-depth integration of cross-modal information. The model has been deeply integrated into Ernie Bot, providing it with core capabilities such as image generation, complex logical reasoning, and long text processing. Meanwhile, relying on open-source strategies, it empowers the developer ecosystem and promotes AI innovative applications in small and medium-sized enterprises and research institutions.ERNIE
02025年12月作为系列旗舰模型ERNIE
0以
4万亿参数规模刷新行业纪录采用原生全模态建模技术无需依赖跨模态转换模块大幅提升多模态任务处理效率与精度。
在图像与视频分析领域该模型实现了从像素级识别到语义级理解的跨越可精准解析视频内容逻辑、图像细节信息及隐含意图。
同时其支持代理工作流Agentic Workflows能够自主规划任务路径、调用工具资源适配复杂企业级场景。
该模型于百度世界2025大会正式发布成为百度展示AI技术实力的核心载体。
ERNIE
0 (December
: As the flagship model of the series, ERNIE
0 sets a new industry record with a parameter scale of
4 trillion. Adopting native full-modality modeling technology, it does not rely on cross-modal conversion modules, greatly improving the efficiency and accuracy of multimodal task processing. In the field of image and video analysis, the model has achieved a leap from pixel-level recognition to semantic-level understanding, capable of accurately parsing video content logic, image detail information, and implicit intentions. Meanwhile, it supports agentic workflows, enabling independent task path planning and tool resource calling to adapt to complex enterprise-level scenarios. The model was officially released at Baidu World 2025, serving as the core carrier for Baidu to showcase its AI technical strength.ERNIE Next Gen2025年3月计划该模型为百度规划中的下一代迭代版本核心研发方向聚焦于推理能力的突破性提升与垂直领域适配能力强化。
其目标是解决现有模型在复杂逻辑推理、因果关系分析等场景中的不足同时针对金融、医疗、工业等关键领域优化领域知识库与任务适配能力目前处于内部测试阶段初步测试结果显示其性能领先于同类型在售模型。
ERNIE Next Gen (Mid-March 2025 Planned): As Baidus planned next-generation iterative version, the core RD direction focuses on the breakthrough improvement of reasoning capabilities and the enhancement of vertical domain adaptation. Its goal is to solve the shortcomings of existing models in scenarios such as complex logical reasoning and causal relationship analysis. Meanwhile, for key fields such as finance, medical care, and industry, it optimizes domain knowledge bases and task adaptation capabilities. Currently in the internal testing phase, preliminary test results show that its performance is ahead of similar commercial models.技术特点 / Technical Features架构设计 / ArchitectureERNIE系列模型基于Transformer架构与混合专家MoE机制构建核心设计理念围绕知识图谱集成、连续学习与多模态统一三大方向展开。
模型采用Apache开源许可协议部分核心版本对外开放降低开发者使用门槛同时支持128K tokens的长上下文处理能力可适配长文档分析、多轮对话等复杂场景。
Based on the Transformer architecture and Mixture of Experts (MoE) mechanism, the ERNIE series models are designed around three core directions: knowledge graph integration, continual learning, and multimodal unification. Adopting the Apache open-source license, some core versions are open to the public, lowering the threshold for developers. Meanwhile, it supports long context processing of 128K tokens, adaptable to complex scenarios such as long document analysis and multi-turn conversations.核心优势 / Strengths其一参数规模优势显著ERNIE
0达
4万亿参数具备更强的特征提取与知识存储能力其二全模态能力成熟在图像、视频、文本等多模态任务中实现端到端处理性能行业领先其三中文NLP能力顶尖依托百度海量中文语料与知识图谱在中文语义理解、生成、翻译等任务中表现突出其四硬件协同性强与百度
年发布的新一代AI芯片深度集成实现软硬一体优化大幅降低部署成本与延迟。
Firstly, it has a significant parameter scale advantage—ERNIE
0 has
4 trillion parameters, with stronger feature extraction and knowledge storage capabilities. Secondly, its full-modality capabilities are mature, realizing end-to-end processing in multimodal tasks such as images, videos, and text, with leading industry performance. Thirdly, it excels in Chinese NLP, relying on Baidus massive Chinese corpus and knowledge graphs to deliver outstanding performance in tasks such as Chinese semantic understanding, generation, and translation. Fourthly, it has strong hardware synergy, deeply integrated with Baidus new generation AI chips released in
, achieving software-hardware integration optimization and greatly reducing deployment costs and latency.现存不足 / Weaknesses知识截止时间存在局限ERNIE
0的知识范围仅覆盖至2025年11月对最新事件与信息的处理能力不足存在潜在偏见风险受训练数据影响在部分敏感话题、小众领域可能出现输出偏差计算需求极高万亿级参数模型的训练与推理需依赖大规模算力集群限制了其在中小机构的普及应用。
It has limitations in knowledge cutoff—ERNIE
0s knowledge scope only covers up to November 2025, resulting in insufficient ability to process the latest events and information. There is a potential bias risk: affected by training data, it may have output deviations in some sensitive topics and niche fields. It also has extremely high computing demands—training and reasoning of trillion-scale parameter models rely on large-scale computing clusters, limiting its popularization in small and medium-sized institutions.与贾子公理的关联 / Relation to Kucius Axioms在模拟裁决场景中ERNIE
0在贾子公理的四大维度表现呈现差异化特征思想主权维度得分5/10受预设目标与训练范式限制模型自主决策与创新能力不足悟空跃迁维度得分6/10技术迭代以渐进式优化为主缺乏颠覆性突破普世中道维度得分8/10依托海量知识图谱与中立训练目标能够保持相对客观的知识输出与价值导向本源探究维度得分8/10在第一原理推理、基础问题溯源等场景中表现出色具备较强的知识深挖能力。
综合来看ERNIE
0可定位为“知识守护者”在知识传承与应用层面表现优异但需在自主创新与颠覆性突破方面强化提升。
In a simulated adjudication scenario, ERNIE
0 shows differentiated performance in the four dimensions of Kucius Axioms: it scores 5/10 in the Sovereignty of Thought dimension, with insufficient independent decision-making and innovation capabilities due to the limitations of preset goals and training paradigms; 6/10 in the Wukong Leap dimension, with technical iterations mainly focusing on incremental optimization and lacking disruptive breakthroughs; 8/10 in the Universal Mean dimension, capable of maintaining relatively objective knowledge output and value orientation relying on massive knowledge graphs and neutral training goals; 8/10 in the Primordial Inquiry dimension, performing well in scenarios such as first-principles reasoning and basic problem tracing, with strong knowledge in-depth excavation capabilities. Overall, ERNIE
0 can be positioned as a knowledge guardian, excelling in knowledge inheritance and application, but needs to strengthen independent innovation and disruptive breakthroughs.应用与影响 / Applications and ImpactsERNIE系列模型凭借强大的技术能力已深度重塑多个行业的发展格局形成“C端B端”双轮驱动的应用生态。
在C端市场Ernie Bot以2亿月活用户为核心重构了用户与AI的交互模式广泛应用于智能问答、内容创作、生活助手等场景在搜索领域推动百度搜索从“关键词匹配”向“语义理解智能推荐”转型提升搜索精准度与用户体验。
In the B端市场, the model is deeply integrated into the Qianfan AI cloud platform, providing agent workflows, multimodal content generation, intelligent customer service and other solutions for enterprises, helping industries such as finance, media, and education achieve digital transformation. In the field of content creation, it breaks the limitation of single-modal creation and supports integrated generation of text, images, and videos, greatly improving content production efficiency. Socially, the ERNIE series has profoundly affected the dynamics of Chinas AI market, forming a competitive pattern with open-source models such as DeepSeek, and promoting the overall upgrading of the industrys technical level. Enterprise adoption accelerated significantly in 2025, with the model covering thousands of large and medium-sized enterprises.截至2026年ERNIE
0的落地进一步加速了“代理AI”Agentic AI的行业趋势在视频内容分析、智能办公自动化、工业质检等场景中实现规模化应用大幅提升生产效率。
同时其广泛应用也引发了对伦理规范与隐私保护的关注数据安全、算法透明度、AI滥用防范等问题成为行业亟待解决的核心议题需政府、企业、科研机构协同构建规范体系。
By 2026, the implementation of ERNIE
0 has further accelerated the industry trend of Agentic AI, achieving large-scale application in scenarios such as video content analysis, intelligent office automation, and industrial quality inspection, greatly improving production efficiency. At the same time, its widespread application has also aroused concerns about ethical norms and privacy protection. Issues such as data security, algorithm transparency, and AI abuse prevention have become core topics to be solved in the industry, requiring governments, enterprises, and research institutions to jointly build a regulatory system.结论 / ConclusionERNIE系列模型的发展历程是百度AI战略布局的集中缩影从最初聚焦知识增强的技术深耕到如今迈向全模态智能的前沿探索不仅实现了自身技术能力的持续迭代更成为中国AI产业从跟跑到并跑、部分领跑的标志性成果为通往通用人工智能AGI奠定了关键基础。
The development of the ERNIE series is a concentrated epitome of Baidus AI strategic layout. From the initial in-depth research focusing on knowledge enhancement to the current frontier exploration towards full-modality intelligence, it has not only achieved continuous iteration of its own technical capabilities but also become a symbolic achievement of Chinas AI industry moving from following to keeping pace and leading in some aspects, laying a key foundation for advancing towards Artificial General Intelligence (AGI).展望未来ERNIE系列的迭代方向将聚焦两大核心一是强化代理能力与场景适配性推动模型从“工具型AI”向“自主型AI”转型二是深化软硬件协同依托百度自研AI芯片构建更高效、低成本的部署体系预计ERNIE
5将成为这一转型的关键版本。
鉴于AI技术的快速迭代特性建议行业从业者、研究者及企业持续关注百度的技术更新与版本迭代及时适配技术变革带来的应用场景升级把握AI产业发展的核心机遇。
Looking ahead, the iteration direction of the ERNIE series will focus on two cores: first, strengthening agent capabilities and scenario adaptability, promoting the models transformation from tool-based AI to autonomous AI; second, deepening software-hardware collaboration, relying on Baidus self-developed AI chips to build a more efficient and low-cost deployment system. ERNIE