ERNIE Bot

百度launched 的知识增强large language model，智能伙伴 and AI assistant

large language model

https://yiyan.baidu.com/

Visit Website

In-depth Report

Wen Xin Yi Yan (English name: ERNIE Bot) is a new generation of knowledge-enhanced large language model developed by Baidu. It is also the core product of Baidu's generative AI products. Wen Xin Yi Yan is not only the user's smart partner, who can chat with you, answer questions, and draw pictures; it is also the user's AI assistant, which can provide inspiration, write copy, read documents, and intelligently translate, helping users complete work and study tasks efficiently. Wenxinyiyan was officially released on March 16, 2023. It is the first large language model product in China to benchmark ChatGPT. In November 2025, Baidu released Wenxin Big Model 5.0, achieving a comprehensive upgrade from single dialogue to multi-modal intelligence. In March 2026, the application was officially renamed "Wenxin" and returned to its original brand name. Wen Xin Yi Yan is built on the ERNIE large model technology independently developed by Baidu. It relies on Baidu's profound accumulation of more than ten years of search and has 550 billion factual knowledge graphs and trillions of web page data as the training basis. As one of the few AI products in the world with a four-layer full-stack technology layout of "chip layer, framework layer, model layer, and application layer", Wen Xinyiyan has natural advantages in Chinese understanding and cultural adaptation. At present, Wenxin Yiyan has attracted more than 650 companies to announce their access to its ecosystem, gradually forming an enterprise application ecosystem with the Wenxin large model as the core. In the 2024 China Large Model Evaluation Strength Ranking, Wen Xinyiyan ranked among the top three.
Baidu, as China's leading Internet technology company, launched ERNIE version 1.0 as early as 2019, and is one of the first domestic technology companies to deploy large language models. After years of technology iteration and data accumulation, Baidu has established a complete AI technology system. Baidu full-stack technology architecture: Globally, Baidu is one of the few technology companies with a four-layer full-stack technology layout. The first layer is the chip layer. Baidu has independently developed the Kunlun Core series of AI chips; the second layer is the framework layer, and Baidu has open sourced the Feipiao deep learning platform; the third layer is the model layer, and Baidu has built the Wenxin large model family with ERNIE as the core; and the fourth layer is the application layer, where Baidu empowers core businesses such as search, intelligent cloud, and autonomous driving with AI capabilities. Data advantage: As the world's largest Chinese search engine, Baidu Search processes billions of search requests every day and has accumulated massive Chinese corpus data. Wen Xinyiyan's training data includes: trillions of web page data, billions of search data and image data, tens of billions of daily voice call data, and 550 billion factual knowledge graphs. These data resources provide Wen Xin Yi Yan with a unique Chinese language environment training basis. Product development history: On March 16, 2023, Baidu founder Robin Li personally released Wen Xinyiyan on the platform, and the first batch of users began to experience the product through invitation test codes. On November 1, 2023, Wenxin Yiyan officially launched a membership model. After activation, Wenxin Large Model 4.0 capabilities can be unlocked, becoming the first large model product in China to charge C-end users. In September 2024, Wen Xin Yi Yan 4.0 version was renamed "Wen Xiao Yan" when it was launched to expand more AI capabilities. On February 13, 2025, Baidu announced that the basic and professional versions of Wenxin Yiyan are fully free and open, and you can use all text generation, AI painting, code writing and other functions without paying. On June 30, 2025, Baidu made a landmark decision to officially open source the Wenxin large model, becoming an important force in China's large model open source ecosystem. In November 2025, Wenxin large model version 5.0 was released, with performance improved by nearly 80% compared to the previous version. In March 2026, the application was renamed "Wenxin" again, officially returning to the original brand name. Technical architecture: Wenxinyiyan adopts the new generation of knowledge-enhanced large language model technology independently developed by Baidu. The core technology includes six key technologies: supervised fine-tuning, human feedback reinforcement learning (RLHF), prompt engineering, knowledge enhancement, retrieval enhancement, and dialogue enhancement.
Wenxinyiyan provides a wealth of AI functions, covering multiple scenarios such as dialogue Q&A, content creation, document processing, and multi-modal generation. Dialogue question and answer functions: intelligent chat, question and answer, semantic understanding, AI question and answer. Users can have natural language conversations with Wen Xinyiyan, get answers to various questions, support multiple rounds of conversation context understanding, and be able to remember previous conversation content and continue communicating on this basis. Content creation functions: AI writing, copywriting, intelligent expansion, and article interpretation. Wenxinyiyan can generate various types of text content according to user needs, including but not limited to marketing copywriting, press releases, story creation, poetry generation, etc. Especially in the creation of Chinese content, Wen Xinyiyan has demonstrated strong cultural understanding. Document processing functions: reading documents, intelligent document management, and document generation. Users can upload documents for Wenxinyiyan to analyze, summarize, and translate, or Wenxinyiyan can generate various documents according to needs. Multi-modal functions: drawing and recognizing pictures, image generation, Vincent video, AI speech synthesis. Wenxinyiyan supports generating images based on text descriptions, and can also identify and interpret image content. Wenxin 5.0, released in November 2025, further strengthens multi-modal capabilities and supports native full-modal understanding and generation. Translation language functions: intelligent translation, machine translation, cross-language understanding. Wenxinyiyan supports translation in multiple languages and can help users with cross-language communication and document translation. Office auxiliary functions: AI office, chart design, data processing. Wen Xin Yi Yan can help users handle various office tasks and improve work efficiency. Five core scene experiences: According to Baidu’s official release and user testing, Wen Xin Yiyan’s performance in the following five core scenes: literary creation scene, business copywriting creation scene, mathematical calculation scene, Chinese understanding scene, and multi-modal generation scene.
Wen Xin Yi Yan adopts a freemium (free value-added) business model. Basic functions are free to use, while advanced functions require membership. Membership mode: The monthly fee for professional version membership is 59.9 yuan/month, and the continuous monthly discount is 49.9 yuan/month. After activation, all capabilities of Wenxin Large Model 4.0 can be unlocked. Starting from February 13, 2025, Baidu further expanded the scope of free rights and announced that the basic and professional versions of Wenxin Yiyan will be fully free and open, and all functions (including text generation, AI painting, code writing, etc.) can be used without paying. Enterprise-level services: In addition to C-side services for individual users, Wen Xinyiyan also provides B-side services for corporate customers. API interface service: Baidu Smart Cloud opens Wenxin Big Model API interface to enterprise customers, and enterprises can carry out secondary development and application integration according to their own needs. Qianfan Large Model Platform: Baidu Qianfan Large Model Platform provides API services of Wenxin Large Model. Enterprise users can easily access Wenxin Large Model capabilities and build their own AI applications. Business model analysis: Wen Xinyiyan’s charging model has the characteristics of differentiated stratification. The free version is used to attract user experience, and the paid version provides more powerful model capabilities to form a clear user stratification. Prominent value: Baidu emphasizes that Wen Xinyiyan’s charging model can create actual value for users, and its membership pricing strategy is reasonable.
Positive comments: The response speed is fast, which is one of the advantages generally recognized by users. He has strong business copywriting skills. Many users say that Wenxinyiyan is excellent in creating business copywriting. Especially when there is no inspiration, it can generate content as a reference, and the effect is very good when combined with its own characteristics. Chinese understanding is okay. Although there is still a gap compared with GPT-3, Wen Xinyiyan has shown good performance in Chinese understanding. The picture generation effect is good, especially in the generation of Chinese-style pictures, and it can generate images with Chinese traditional cultural characteristics. There is a foundation for technological accumulation. Users believe that Baidu, as the first domestic team to follow up on large models, has demonstrated its technological accumulation and capabilities. Negative comments: The depth of semantic understanding is insufficient. Some users reported that Wen Xinyiyan did not understand deeply enough when dealing with some complex semantics, and sometimes deviations in literal meaning understanding occurred. Mathematical reasoning ability is weak. Logical reasoning ability is a common problem in current large AI models. Wen Xinyiyan still has room for improvement in mathematical reasoning. The intelligence of answers is limited. Some users think that Wen Xinyiyan’s answers are sometimes not intelligent enough and there is still a gap compared with ChatGPT. There is a gap between Wen Xinyiyan and ChatGPT. Many users compare Wen Xinyiyan with ChatGPT and think that there are still gaps in some aspects.
Professional evaluation performance: According to the authoritative evaluation data of SuperBench 2024 and SuperCLUE 2025, in the comparison of the Chinese capabilities of the four mainstream models: Wenxin Yiyan 4.0, GPT-4o, Claude 3, and Gemini 2.5, Wenxin Yiyan performed well in terms of Chinese core capabilities, and each has its own winner or loser with the international mainstream models. According to a detailed evaluation report released by Baidu Developer Center, Wenxinyiyan 4.5 has achieved breakthroughs in the following aspects: text generation speed is increased by 47% compared with the previous generation, single inference delay is compressed to less than 320 milliseconds, throughput is increased by nearly 2 times, image-text association accuracy is increased from 68% to 89%, and new knowledge accuracy is 91%. According to evaluation data, Wenxin 5.0 has improved by nearly 80% compared to the previous version, and has the same performance as the popular star model MiniMax M2 in the same period. In the text arena of LMSYS Chatbot Arena, Wenxin 5.0 Preview ranked at the top with an Elo score of 1432 points. According to the "2024 China Large Model Evaluation Strength Ranking TOP20" released by iiMedia Research, Wen Xinyiyan ranks among the top three, and together with Doubao, Kimi, etc., form the first echelon of China's large models. Comparison with competing products: Compared with ChatGPT, make a comprehensive comparison in terms of technical architecture, training data, application scenarios, etc. Both have their own advantages. Domestic large-scale models have shifted from "technical competition" to deep scene development: Tongyi wins in ecology, DeepSeek is stronger in reasoning, Beanbao is unbeatable in cost performance, and Kimi focuses on long texts. It is recommended that users choose appropriate tools based on actual usage scenarios.
Copyright risk: Source of AI video material. There are reports that Wen Xinyiyan’s AI image-to-text video conversion function has copyright risks. The source of Baidu video material is unknown, which may cause the creator to get into copyright disputes. Untagged content. Some users reported that some of the articles generated by Wen Xinyiyan were not tagged, and there may be copyright or signature issues. Security risk: Ethical boundary test. Some testers conducted an "extreme questioning" test on Wen Xinyiyan. The results showed that Wenxinyiyan has a corresponding safety mechanism, but it still needs to be continuously optimized. Content review. As a public-facing AI product, Wen Xinyiyan needs to comply with relevant laws, regulations and content security requirements. Access restrictions: Some users reported that access to Wen Xinyiyan was restricted. This may be due to technical problems or policy adjustments. Market competition risk: With the rise of large domestic models such as Doubao, Kimi, and DeepSeek, Wen Xinyiyan faces increasingly fierce market competition. International gap: Compared with leading international models such as GPT-4, Claude 3, etc., Wen Xinyiyan still lags behind in some technical indicators.
Applicable people: Content creators, copywriters, self-media people, writers and other text creators can use Wenxin Yiyan to assist creation, get inspiration and generate first drafts. Office workers, professionals who need to process various documents, reports, and emails can use Wen Xin Yi Yan to improve document processing efficiency. Student groups, students who need to write papers, organize materials, and learn knowledge can use Wen Xin Yi Yan to assist their learning. Developers who need to develop AI applications can access Wenxin Large Model API through Baidu Qianfan Large Model Platform for secondary development and application integration. Ordinary users who want to experience AI dialogue and intelligent assistant functions can use Wenxinyiyan for daily Q&A, knowledge query, language translation and other tasks. Usage suggestions: Clarify the task type, and different tasks are suitable for different usage methods. Step-by-step processing. Since there is a limit of 1024 words for a single input, it is recommended to process it in sections and gradually improve it when doing long text tasks. Manual review. It is recommended that the content generated by Wen Xinyiyan be reviewed manually. In combination with other tools, multiple AI tools can be used in combination according to different needs. Pay attention to version updates. Wen Xinyiyan maintains a high iteration speed. It is recommended to pay attention to version updates. Free priority is given. At present, the basic version and professional version are fully open for free. Ordinary users can experience all functions for free first.
Product advantages: As one of the first major language model products launched by Baidu in China, Wenxinyiyan has the following core advantages: strong Chinese understanding ability. Relying on the massive Chinese corpus accumulated by Baidu search, Wenxinyiyan has natural advantages in Chinese understanding and localization adaptation. The technical layout is complete. Baidu's four-layer full-stack technology layout provides Wen Xinyiyan with complete technical support from chips to applications. The ecosystem is complete, with more than 650 companies connected to the ecosystem, forming a relatively complete enterprise application ecosystem. Continuous iteration, from 3.5 to 4.0 to 4.5 to 5.0, maintained a high product iteration speed. Localized services, as domestic products, are more suitable for the needs of Chinese users in terms of compliance requirements and service response. Insufficient products: There is a gap between Wen Xinyiyan and internationally leading products. Compared with internationally leading products such as GPT-4 and Claude 3, Wen Xinyiyan still lags behind in some technical indicators. Reasoning ability needs to be improved. Logical reasoning ability is a common problem in current large AI models. Wen Xinyiyan still has room for improvement in this regard. Long text processing challenges, attention may be distracted when processing extremely long text, and it needs to be processed step by step. Development trend: Technology continues to upgrade. Wenxin 5.0 has achieved native full-mode upgrades and greatly improved performance. Open source ecological construction. In June 2025, Baidu announced the open source Wenxin large model, which will further promote the development of China's large model open source ecosystem. Application scenarios are expanding. With technological advancement and ecological improvement, Wen Xinyiyan will play a role in more industries and application scenarios. Free strategy expansion, the basic and professional versions in February 2025 will be completely free, showing Baidu’s strategy of promoting ecological development by expanding the user base. Taken together, Wenxinyiyan is an important representative of China's large language model products and has obvious advantages in Chinese processing and localization services.

User Reviews

JMurphyX

—

搜索这块文心是真的能打，国产第一没跑。

Susan_Chavez_99

—

放心写这个功能真是被低估了，我们单位天天写公文报告，以前用别家AI得逐句核实事实，现在选个发言稿体裁把主题输进去，它基于权威数据源生成还自动做事实校验，出来的东西基本能直接交，体制内的兄弟们真该试试。

AXfra

—

6月25号百度终于把文心一言网页版、文心App网页版、百度文心助手三合一了，早该这么干，之前入口太散，想聊天去一个地方想做PPT又是另一个地方。合并后是清爽了，但功能堆太多，各个功能之间衔接还是不够顺，跟夸克那种全链路体验比差点意思。

Judy.Brooks_2022868

—

查最新资讯是真方便，问它7月最新AI行业动态，它不是随便搜几条拼一起，而是筛权威来源、剔除冗余虚假信息，最后给你一份条理清晰的整合回答，末尾还带信源引用。对比下来DeepSeek的联网偏给你链接自己看，豆包搜出来偶尔混营销内容，文心是那种搜完直接帮你整理好能用的类型。

Jacqueline86

—

编程是真拉胯，让它做个小游戏，几百行代码生成完界面全是bug，根本没法玩，这块还是老老实实用DeepSeek吧。

Alice_Torres520717

—

文心任务这个一键托管挺有意思，你说完需求就能去干别的，它自己跑完给你交付H5、PPT、Word。我让它做个AIGC行业调研H5白皮书，出来三大章节配数据卡片图表，排版都统一好了。就是局部想微调不太灵活，适合快速出初稿，精修还得自己上手。

PBrown_2021

—

全模态是文心相比DeepSeek和Kimi的明显优势，看图说话、视频理解、图表解读都能干，Kimi连图都画不了。

YIjen

—

用文心一言有段时间了，中文语境理解很贴合，日常写文案、整理工作笔记不用复杂指令就能出结果，知识库更新也及时，查行业基础问题很快。但创意生成太弱，写故事编脚本全是模板套路，这点挺劝退的。

TheNicholasHernes_88

—

客观评价，文心其实就是个超级增强版百度搜索，查政策、找实时新闻、写四平八稳的公关稿很稳，本土知识图谱和时效性基本无可匹敌。但你让它写有灵气的创意策划或者解高难度数理题，就露怯了，输出刻板冗杂，还老强行把你拉回它那套安全模板。

ArthurFisher_2023

—

数学是真不行，稍微难点的推导就翻车。

邹红

—

说点不好的，文心一言独立App用户流失挺严重的，听说月活一年差不多腰斩了，虽然嵌在百度App里的文心助手月活过两亿，但这恰恰说明它自己立不住，太依赖百度搜索导流，用户用完就走留不住人。

BlockFi88

—

深入研究功能拿来做行业调研报告不错，自动多轮检索加结构化输出，省了我自己一条条筛信息的时间。

KAlvarez_2022

—

改名改得我都晕了，文心一言文心文小言百度文心到底哪个是哪个啊。

Melissa367_q

—

深度测评下来，文心适合职场新人快速出初稿，文案、PPT大纲这些办公场景表现挺亮眼，但复杂逻辑推理和长文本还是有提升空间，深度创作还得人工优化把关。

CarlBaker

—

用文心合并三份会议纪要成周报提纲，五分钟出稿效率提了三倍。它长文档处理挺强的，180页PDF加Excel扔进去，让它给我5条风险点还标注条款，三分钟就定位好了。中文写作加搜索增强这个组合确实实用，就是重要数据还是得人工复核一遍。

MeganAustin

—

看了下第三方评测，说文心5.1在MMLU-Pro那种广博知识测试里排末位，coding也不稳定，但搜索榜和数学工具调用是真上去了，评价挺两极的。

烟雨498

—

学生党实测，写作文演讲稿总结笔记都很快，结构清晰语句通顺，还能按要求调文风，遇到难题提问讲解步骤也很详细，比单纯看答案更容易理解知识点。依托百度知识图谱，它对成语常识专业概念的理解很到位，很少出现生硬的机翻腔，免费版够日常用了，稳定性强出错率也低。

桃花915

—

海外用起来挺折腾的，注册要大陆手机号，付费也麻烦，英文任务不如ChatGPT和Claude，但如果你就是要中文原生的理解力和百度生态整合，那它就是默认选项。

Theresa.Hernandez_2020

—

高阶表格操作别指望它，简单的业绩统计求均值筛选没问题，一到复杂数据透视、多表联动就掉链子，落后Claude和Gemini一大截。

NWilson

—

客观说文心5.1性价比是真高，官方说预训练成本只有同规模模型的6%，API价格便宜一大截，做agent开发的话国产里算第一梯队，工具调用也挺稳的。

BruceBerg

—

偶尔会一本正经说胡话，精准数据类的问题容易答非所问，长文本写多了逻辑也会跑偏，衔接生硬。当实用工具还行，别指望它多有灵气。

FrankGreen

—

会员分两档，普通会员49一个月，超级会员99，免费版高峰期偶尔限流提示排队升级。轻度用的话免费完全够，写得多才值得上订阅。

Diana.Hart_X

—

免费还能实时联网，日常够用了，香。

Jason.Reed007

—

创意文案有时候偏正式，得多调几次才有那味儿，部分复杂多模态任务表现也一般，但瑕不掩瑜，免费版应付日常足够了。

廖欣睿

—

文心的英文能力确实一般，我拿它写英文邮件明显不如ChatGPT顺，但中文任务上是真强，古诗词成语这种拿捏得死死的。

JBell369

—

ERNIE在中国生态整合是最深的，百度搜索、地图、网盘都打通了，Kimi强在超长上下文，Qwen胜在多语言和开源，各有各的活法吧。

BruceKelly

—

中文AI里它是真省心，回答很贴国内的表达习惯。

DScott596

—

文心一言的联网搜索太香了，生成的文章有实时新闻支撑，效率拉满。

董红刚

—

百度生态深度整合yyds！搜索+AI一体化，其他竞品真学不来。

MButler_2024

—

深度搜索功能上线了，体验了一下，确实比普通模式更专业，适合做研究。

ERNIE Bot

In-depth Report

User Reviews

Similar Products