BIDU(BIDU)
Search documents
2.4万亿参数原生全模态大模型 文心5.0正式版上线
Sou Hu Cai Jing· 2026-01-22 18:45
吴甜现场介绍了"文心导师"计划的最新进展。文心导师队伍持续壮大,目前已吸纳835位来自科技、金融、文化、教育、医疗、能源等十余个重点行业以及 数理化生文史哲等学科的专家,在知识传授、鉴赏评价、专业校准方面对大模型进行指导,帮助模型在逻辑严谨性、专业深度、创意质量和价值观对齐等方 面不断精进。 在40余项权威基准的综合评测中,文心 5.0 正式版的语言与多模态理解能力超越Gemini-2.5-Pro、GPT-5-High等模型,稳居国际第一梯队。图像与视频生成 能力与垂直领域专精模型相当,整体处于全球领先水平。 有分析认为,文心5.0正式版的上线,意味着原生全模态技术路线逐渐走向成熟与实用,体现了中国模型厂商在多模态大模型底层自主创新能力,有助于进 一提升中国AI在全球AI产业竞争中的技术话语权。 据了解,2025年11月,百度发布文心大模型5.0并上线Preview版本。此后,文心5.0系列模型多次登上LMArena全球大模型竞技场,多次位居文本榜和视觉理 解榜国内第一,跻身国际第一梯队。1月15日,文心5.0以1460分位列LMArena文本榜国内第一、全球第八,超过GPT-5.1-High、Gemini- ...
当AI开始“插话” 文心内测群聊功能
Bei Jing Shang Bao· 2026-01-22 15:54
比达分析师李锦清向北京商报记者表示:"此群聊非彼群聊。可以这样理解,文心App的群聊以结果为 目的,即时通讯App的群聊以社交为目的。" 在与媒体沟通时,百度内部人士表示:"我们没有考虑过要做一个微信或者取代微信,这个不现实。百 度这一产品是基于需求和任务出发的。大家如果有特定的任务,比如家庭健康、小组作业、旅游规划, 或者是一些特别复杂的决策,我们希望提供一个平台给大家,大家可以尝试在这里跟同学、朋友一起, 借助AI的能力,更好地完成任务。" 以"文心健康管家"为例,当群聊助手识别到讨论内容涉及专业医疗健康问题时,会自动邀请健康领域的 垂直类智能体入群解答。 按照百度内部人士的解读,该产品旨在解决传统群聊中信息过载、结论稀缺及协作失效等痛点。致力于 将群聊从单纯的交流工具升级为协作空间,让群聊成为产出结果、推动行动的场所。形象地说,文心群 聊就像是把一家外包公司拉进了群,通过AI技术辅助用户完成从讨论到执行的全过程。 一对一交互不是Agent(智能体)的上限。1月22日,北京商报记者获悉,文心App正在内测多人、多 Agent群聊功能。该功能支持在同一群聊中调动多个AI角色,AI能根据讨论氛围主动应答。目前 ...
2.4万亿参数原生全模态大模型 百度文心5.0正式版上线
Zheng Quan Shi Bao Wang· 2026-01-22 14:06
1月22日,百度发布并上线原生全模态大模型文心5.0正式版。该模型参数达2.4万亿,采用原生全模态统 一建模技术,具备全模态理解与生成能力,支持文本、图像、音频、视频等多种信息的输入与输出。目 前,个人用户可在文心APP、文心一言官网体验,企业与开发者可通过百度千帆平台进行调用。 百度集团副总裁、深度学习技术及应用国家工程研究中心副主任吴甜介绍,与业界多数采用"后期融 合"的多模态方案不同,文心5.0的技术路线采用统一的自回归架构进行原生全模态建模,将文本、图 像、视频、音频等多源数据在同一模型框架中进行联合训练,使得多模态特征在统一架构下充分融合并 协同优化,实现原生的全模态统一理解与生成。 此外,文心5.0采用超大规模混合专家结构,具备超稀疏激活参数,激活参数比低于3%,在保持模型强 大能力的同时,有效提升推理效率。同时,基于大规模工具环境,合成长程任务轨迹数据,并采用基于 思维链和行动链的端到端多轮强化学习训练,显著提升了模型的智能体和工具调用能力。 得益于原生全模态建模,文心5.0在多模理解、代码、创意写作等能力上实现显著突破。 大会现场,仅输入一段博主复刻"活了么"App教程视频,文心5.0便能自动 ...
【太平洋科技-每日观点&资讯】(2026-01-23)
远峰电子· 2026-01-22 13:00
Market Overview - Major indices showed positive performance with ChiNext Index up by 1.01%, North Star 50 by 0.69%, Shenzhen Component Index by 0.50%, Sci-Tech 50 by 0.41%, and Shanghai Composite Index by 0.14% [1] - TMT sector led the gains with SW Communication Network Equipment and Devices increasing by 4.15%, SW Printed Circuit Board by 3.84%, and SW Horizontal General Software by 3.19% [1] - Conversely, TMT sector also saw declines with SW Integrated Circuit Packaging and Testing down by 3.65%, SW Semiconductor Equipment by 2.28%, and SW Semiconductor Materials by 2.09% [1] Domestic News - BlueTech Optical raised 1.055 billion for AR optical product industrialization, with 500 million allocated for advanced production and testing equipment [2] - Fudan University team successfully created large-scale integrated circuits within flexible polymer fibers, achieving high precision interconnections [2] - Passive component manufacturer Huaxin announced a price increase for resistors effective February 1, aligning with industry leader Yageo's plans [2] - Yuexin Semiconductor's fourth-phase project commenced, with a total investment of approximately 25.2 billion for a production line capable of 40,000 12-inch wafers per month [2] Overseas News - Kioxia's storage division reported that its 2026 production capacity is sold out, and it plans to enhance factory efficiency to meet demand [3] - Cadence partnered with Microsoft to develop a high-performance, low-power memory system for data centers, set for deployment in Microsoft's facilities [3] - The EU Commission proposed a draft revision of the Cybersecurity Law, aiming to phase out high-risk suppliers in 18 critical infrastructure sectors [3] - Display industry trends indicate that panel manufacturers are adjusting utilization rates to balance supply and demand, with expectations for price increases in 2026 [3] AI News - Adobe introduced new AI features in Acrobat for document review and content creation, enhancing workflow efficiency [4] - Baidu launched the Wenxin large model 5.0, featuring 2.4 trillion parameters and advanced multimodal capabilities [4] - Tencent Cloud released CodeBuddy 2.0, enhancing AI application development capabilities for enterprises [4] - Microsoft Research introduced the Rho-alpha AI model, capable of interpreting natural language commands for complex robotic tasks [4] Industry Tracking - China's commercial space sector is projected to complete 50 launches in 2025, accounting for 54% of total launches, with 311 commercial satellites entering orbit [5] - A high-throughput flexible brain-computer interface system was developed, achieving real-time Chinese language decoding [5] - A pilot production line for 6N-grade synthetic quartz sand was established in Liaoning with an investment of 3.36 million [5] - UBTECH Robotics signed a cooperation agreement with Airbus to supply robotic products for aerospace manufacturing [5] Earnings Forecasts - Yidao Information expects a net profit of 60 million to 78 million for 2025, a year-on-year increase of 76.04% to 128.85% [6] - Zhaoyi Innovation anticipates a net profit of 1.61 billion for 2025, up 46% year-on-year [6] - Ruichuang Micro expects a net profit of 1.1 billion for 2025, reflecting a 93% increase [6] - Founder Technology forecasts a net profit of 430 million to 510 million for 2025, a growth of 67.06% to 98.14% [6] High-Frequency Data Updates - As of January 22, 2023, international DRAM spot prices showed varied trends, with DDR5 16G averaging $36.667, unchanged from the previous day [7] - Semiconductor material prices were updated, with 4N zinc oxide powder priced at 1,495 yuan per kilogram, reflecting a daily increase of 10 yuan [7]
百度文心大模型5.0正式发布 千帆平台累计开发Agent超130万个
Xin Hua Cai Jing· 2026-01-22 12:41
Core Insights - Baidu officially launched the Wenxin large model version 5.0, featuring 2.4 trillion parameters and utilizing a native multimodal unified modeling technology that supports various forms of information input and output, including text, images, audio, and video [2] Group 1: Model Features - Wenxin 5.0 employs a unified autoregressive architecture for native multimodal modeling, allowing for joint training of text, images, video, and audio within the same model framework, enhancing the integration and optimization of multimodal features [2] - Despite having a high parameter count, Baidu implemented a sparse activation strategy, activating less than 3% of parameters to address inference cost issues associated with large models [2] - Official tests indicate that Wenxin 5.0's language and multimodal understanding capabilities are on par with leading international models such as Gemini-2.5-Pro and GPT-5-High [2] Group 2: Application and Development - Baidu has developed a matrix model and specialized models based on the Wenxin foundational large model, with the matrix model aimed at rapid deployment in product-level applications and general scenarios, including Wenxin Lite, video, and voice models [3] - Specialized models target industry applications and vertical scenarios, including search lightning models, e-commerce steam engine models, Wenxin digital human models, and industry-specific large models [3] - The Baidu Qianfan platform has developed over 1.3 million agents, indicating a significant push towards the industrial application of large models [4] Group 3: Industry Context - The global AI industry is entering a new fast track after several years of rapid development, with chatbots primarily based on dialogue or text input remaining the mainstream form in AI applications [4] - Baidu aims to support various intelligent applications with its models and continues to explore AI solutions that empower industries [4]
这场AI市集,大模型拉满你的情绪价值
Yang Zi Wan Bao Wang· 2026-01-22 12:22
Core Insights - The article highlights the launch of Baidu's Wenxin 5.0, a new multimodal AI model with 2.4 trillion parameters, emphasizing its capabilities in understanding and generating various types of information including text, images, audio, and video [3] Group 1: Product Launch and Features - Baidu's Wenxin 5.0 was officially released at the Wenxin Moment Conference in Shanghai, showcasing its advanced multimodal unified modeling technology [3] - The model supports diverse input and output formats, enhancing its applicability across different media [3] - Personal users can access the model through the Wenxin app and website, while enterprises and developers can utilize it via Baidu's Qianfan platform [3] Group 2: User Engagement and Market Impact - The Wenxin assistant has surpassed 200 million monthly active users, indicating strong market interest and engagement [3] - The focus of Baidu's application model development is on real-world applications, ensuring that the model performs reliably and effectively in practical scenarios [3]
百度大模型“上新”
Shang Hai Zheng Quan Bao· 2026-01-22 12:15
从性能上看,文心5.0正式版的语言与多模态理解能力超越Gemini-2.5-Pro、GPT-5-High等模型,稳居国 际第一梯队。图像与视频生成能力与垂直领域专精模型相当,整体处于全球领先水平。 "和目前业内大多采用的后期融合多模态方案不同,文心5.0的技术路线采用统一的自回归架构进行原生 全模态建模,将文本、图像、视频、音频等多源数据在同一模型框架中进行联合训练。"活动现场,百 度集团副总裁、深度学习技术及应用国家工程研究中心副主任吴甜介绍道,这一策略使得多模态特征在 统一架构下充分融合并协同优化,实现原生的全模态统一理解与生成。 得益于原生全模态建模,文心5.0在多模态理解、代码、创意写作等能力上实现显著突破。此外,文心 5.0采用超大规模混合专家结构,具备超稀疏激活参数,激活参数比低于3%,在保持模型强大能力的同 时,有效提升推理效率。同时,基于大规模工具环境,合成长程任务轨迹数据,并采用基于思维链和行 动链的端到端多轮强化学习训练,显著提升了模型的智能体和工具调用能力。 仅输入一段博主复刻"活了么"App的教程视频,在数分钟内,文心5.0便能自动拆解步骤,理解核心交互 逻辑,并直接生成可运行的前端代 ...
AI日报丨苹果计划下半年将SIRI改造为其首款聊天机器人;百度发布文心大模型5.0正式版
美股研究社· 2026-01-22 11:11
整理 | 美股研究社 在这个快速 变 化的 时代, 人工 智能技术正以前所未有的速度发展,带来了广泛的机会 。 《AI日 报 》致力于挖掘和分析最新的AI概念股公司和市场趋势,为您提供深度的行 业 洞察和 价 值 分析。 A I 快 报 【百度发布文心大模型 5.0 正式版】 1 月 22 日,百度"文心 Moment "大会在上海举行。会上,百度正式发布文心大模型 5.0 正 式版,其基于原生全模态建模,拥有 2.4 万亿参数,支持文本、图像、音频、视频等多种信息 理解输出。此前,文心 5.0 以 1460 分位列 LMArena 文本榜国内第一、全球第八,超过 GPT-5.1-High 、 Gemini-2.5-Pro 等多款国内外主流模型。 【 OpenAI 首席财务官为 ChatGPT 引入广告辩解】 OpenAI 首席财务官萨拉·弗里亚尔( Sarah Friar )在瑞士达沃斯世界经济论坛( WEF ) 的一个小组讨论中为公司在 ChatGPT 中加入广告的行为进行了辩护,称这是实现人工智能( AI )准入"民主化"的一种方式。弗里亚尔表示,公司需要"建立强劲的商业模式",以实现将 AI 带给更广 ...
文心内测群聊功能,当AI开始“插话”
Bei Jing Shang Bao· 2026-01-22 09:58
比达分析师李锦清向北京商报记者表示:"此群聊非彼群聊。可以这样理解,文心App的群聊以结果为目的,即时通讯App的群聊以社交为目的。" 在与媒体沟通时,百度内部人士表示:"我们没有考虑过要做一个微信或者取代微信,这个不现实。百度这一产品是基于需求和任务出发的。大家如果有特 定的任务,比如家庭健康、小组作业、旅游规划,或者是一些特别复杂的决策,我们希望提供一个平台给大家,大家可以尝试在这里跟同学、朋友一起,借 助AI的能力,更好地完成任务。" 一对一交互不是Agent(智能体)的上限。1月22日,北京商报记者获悉,文心App正在内测多人、多Agent群聊功能。该功能支持在同一群聊中调动多个AI 角色,AI能根据讨论氛围主动应答。目前,国内尚无同类产品实现多AI在同一群聊中主动协作。过去,AI答其所问;如今,AI言其所察。从工具到协作 者,边界正在模糊。至于这"插话"能否真正帮人成事,而非制造干扰,还需真实场景验证。 近期百度旗下文心App将启动交互界面改版,新增"多人、多Agent群聊"功能,以更加社交化、人性化的方式进行用户交互。目前,群聊功能已在文心App内 测阶段。 文心是百度旗下全能AI助手,为用户提供 ...
模型参数达2.4万亿,百度发布文心大模型5.0
Guo Ji Jin Rong Bao· 2026-01-22 09:53
基于文心基础大模型,百度构建了矩阵模型和专精模型。矩阵模型面向产品级应用和通用场景快速落地,包括文心Lite模型、视频大模型和语音大模 型;专精模型面向行业应用和垂直场景,包括搜索闪电专精模型、电商蒸汽机模型、文心数字人大模型及行业大模型。 贾磊在现场分享了三项技术突破:基于声音 Token 的端到端合成大模型、5分钟超越真人的直播技术、实时交互数字人技术。 1月22日,百度发布并上线原生全模态大模型文心5.0正式版。该模型参数达2.4万亿,采用原生全模态统一建模技术,具备全模态理解与生成能力,支持 文本、图像、音频、视频等多种信息的输入与输出。 据百度集团副总裁、深度学习技术及应用国家工程研究中心副主任吴甜介绍,与业界多数采用"后期融合"的多模态方案不同,文心5.0的技术路线采用 统一的自回归架构进行原生全模态建模,将文本、图像、视频、音频等多源数据在同一模型框架中进行联合训练,使得多模态特征在统一架构下充分融合并 协同优化,实现原生的全模态统一理解与生成。 此外,文心5.0采用超大规模混合专家结构,具备超稀疏激活参数,激活参数比低于 3%,在保持模型强大能力的同时,有效提升推理效率。同时,基于 大规模工具 ...