[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"$fhdT1Be2iMTQOCjDfqyXtk30kHZJt43uWqZ1lnCywEbc":3},{"code":4,"msg":5,"data":6},200,"操作成功",{"id":7,"title":8,"content":9,"digest":10,"source":10,"coverPath":11,"thumbsCoverPath":12,"isTop":13,"isShow":14,"baseClick":13,"clickCount":15,"createTime":16,"typeId":17,"isNewest":18,"newsInfoTypeRespVo":19,"voiceUrl":22,"voiceSize":23,"taskId":24,"releaseTime":25,"titleEn":26,"contentEn":27,"voiceUrlEn":28,"taskIdEn":29,"voiceSizeEn":30},1415,"蚂蚁、OpenAI、DeepSeek卷疯了！国产最强万亿参数旗舰模型Ling-1T开源","\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002Fc3485400866f4a1fa42488e6504d2867\u002Fc6815124670442f89ecf82c1308173f2~tplv-tt-origin-web_gif.jpeg\" width=\"552\" height=\"undefined\" style=\"display: block; margin: auto;\" class=\"ql-align-center\">\u003Cp class=\"ql-align-center\">\u003Cspan style=\"color: rgb(187, 187, 187);\" class=\"ql-lineHeight-1-75\">（图片来源：拍摄于2025世界人工智能大会蚂蚁展台）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">刚刚，全球最新的万亿参数通用语言模型开源了。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">10月9日消息，蚂蚁集团今天发布并开源万亿参数的通用语言模型 Ling-1T。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">据悉，Ling-1T是蚂蚁百灵大模型Ling 2.0 系列的第一款旗舰模型，也是蚂蚁百灵团队迄今为止推出的规模最大、能力最强的非思考大模型。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">基准测试数据显示，在有限输出Token条件下，Ling-1T于多项复杂推理基准中取得SOTA表现，同时在代码生成、软件开发、竞赛数学、专业数学、逻辑推理等多项高难度基准测试上，Ling-1T均取得领先成绩，多项指标位居开源模型的榜首。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">以竞赛数学榜单 AIME 25为例。Ling-1T以平均4000+ Token的消耗达到了70.42%的准确率，优于Gemini-2.5-Pro，后者平均达5000+Tokens，准确率为70.10%。\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">蚂蚁Ling-1T用更少的Token实现了更高的准确率，性价比优于谷歌Gemini系列，展现出在推理精度和思考效率综合能力上的优势。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002Ff7a223e5db994fd0a88e0a14e946a03e\u002Fa6aae0c2d02d4b465162db43fe4fc3cc~tplv-tt-origin-web_gif.jpeg\" width=\"564\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">不仅是蚂蚁，国庆中秋假期前后，OpenAI、阿里、智谱、DeepSeek等中国和美国的AI大模型加速竞争。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">OpenAI发布AI视频模型Sora2，并在开发者大会上公布GPT-5 Pro、ChatGPT框架Apps SDK等，引发全球关注；\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">DeepSeek则发布DeepSeek-V3.2-Exp，全面适配国产算力卡，并且训练推理提效、API同步降价；\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">阿里通义则发布新一代原生全模态大模型Qwen3-Omni，以及全开源通义DeepResearch模型、框架、方案等；\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">智谱则发布旗舰模型GLM-4.6，并凭借SOTA的Agentic Coding能力，GLM-4.6登顶Hugging Face Trending 全球第一、LMArena 上排名开源第一、全球第四，与Qwen Max Preview（闭源）并列国内第⼀。\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">正如零一万物创始人、CEO李开复所说，底座基础模型是一场千亿级别公司的“军备竞赛”。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">英伟达CEO黄仁勋则指出，\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">过去六个月AI计算需求大幅上升，英伟达新一代架构Blackwell的芯片需求“非常非常高”。他认为这标志着“新一轮工业革命”开始。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">英伟达上月宣布，计划未来十年向OpenAI共计投资1000亿美元，支持OpenAI部署需要10千兆瓦电力的英伟达系统，相当于400万至500万个GPU。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">OpenAI CEO奥尔特曼（Sam Altman）周三表示，\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">技术突破的关键在于更聪明的模型、更长的上下文处理能力和更好的记忆系统。\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">AGI的核心在于发现新知识，AGI 最重要的衡量标准已不是通过某个测试，而是 AI 开始具备“发现新知识”并扩展人类知识边界的能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">国产最强万亿参数旗舰模型来了\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">近年来，大语言模型发展迅速，尤其是DeepSeek热潮，引发学界和业界对通用人工智能（AGI）的广泛讨论，而混合专家（MoE）模型在特定任务中表现优异，但训练依赖高性能计算资源，成本高昂，限制了其在资源受限环境中的应用。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">蚂蚁Ling团队认为，虽然MoE模型训练对高性能AI芯片（如H100和H800）需求大，且资源供应，但低性能加速器更易获取且单位成本效益高，因此，模型需要能在异构计算单元和分布式集群间切换的技术框架。同时在AI Infra部分，在跨集群、跨设备的兼容和可靠层面进行性能优化。该公司设定的目标是“不使用高级GPU”来扩展模型。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">今年3月，蚂蚁集团Ling Team团队利用AI Infra技术开发两个百灵系列开源MoE模型Ling-Lite和Ling-Plus，\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">前者参数规模168亿，Plus基座模型参数规模高达2900亿，引发广泛关注。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">随后，蚂蚁还公布了语音 AI 框架 Ming‑UniAudio、新一代推理模型Ring-flash-2.0等模型产品。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">目前，蚂蚁AGI团队主要由蚂蚁集团副总裁、首席技术官何征宇负责。据悉，何征宇获佐治亚理工学院计算机博士学位；2012年至2018年就职于谷歌，在谷歌创立并领导了开源项目gVisor；2018年何征宇加入蚂蚁集团，负责公司技术基础设施建设，主导了蚂蚁云原生化、绿色计算实践、机密计算创新、开源战略布局等重要项目，而最新的百灵大模型计划聚焦在生活服务、金融服务、医疗健康等场景。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">如今，蚂蚁百灵大模型团队终于公布最新、全球性能较强的万亿参数非思考模型Ling-1T。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Ling-1T沿用Ling 2.0架构，在20T+tokens高质量、高推理浓度的语料上完成预训练，支持最高 128K 上下文窗口，通过“中训练+后训练”的演进式思维链（Evo-CoT）极大提升模型高效思考和精准推理能力。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">值得一提的是，Ling-1T 全程采用 FP8 混合精度训练（部分技术已开源），是目前已知规模最大的使用 FP8 训练的基座模型。这一设计为训练带来了显著的显存节省、更灵活的并行切分策略和 15%+ 的端到端加速。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">强化学习阶段，蚂蚁百灵团队创新性地提出了LPO方法(Linguistics-Unit Policy Optimization，LingPO) ，这是一种以“句子”为粒度的策略优化算法，为万亿参数模型的稳定训练提供了关键支持。另外，蚂蚁百灵团队提出了新的混合奖励机制，在确保代码正确、功能完善的同时持续提升这个万亿基座对视觉美学的认知。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">在ArtifactsBench 前端能力基准上，Ling-1T得分达59.31，并且在可视化和前端开发任务领域以显著优势位居开源模型榜首。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002F6dcda3f485a34882a0737bafc8172d7f\u002F8fa1faf093335c494e0fe416984784f1~tplv-tt-origin-web_gif.jpeg\" width=\"692\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">据蚂蚁百灵团队透露，除了Ling-1T这款非思考模型，蚂蚁百灵团队还在训练万亿参数级的深度思考大模型Ring-1T，已在9月30日开源了preview版。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">目前，开发者都可以体验Ling-1T模型。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">吸金1927亿美元，全球AI投资创纪录\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">据PitchBook统计，今年以来，全球AI领域初创公司已吸引创纪录的1927亿美元风险投资，2025年有望成为首个逾一半风投资金流入AI行业的年份。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">大部分风投资金流向相对成熟的初创公司。其中，OpenAI、Anthropic、xAI等头部AI创业公司本季均募集了上百亿美元，而一些知名度较低的新创公司则举步维艰，尤其是那些并非专注于AI的企业。PitchBook发现，IPO和并购环境紧缩也让部分风险投资人不愿押注未经验证的公司。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">最近一个季度，美国风险投资将62.7%的资金投向AI领域的公司，全球风险投资该比例为53.2%。今年以来，全球风投总额为3668亿美元，其中美国风投达到2502亿美元。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">PitchBook研究主管Kyle Sanford表示，“市场正在分化，要么在做人工智能，要么不是，要么是大公司，要么不是。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">PitchBook AI和网络安全高级研究分析师 Dimitri Zabelin 表示，目前主要的退出趋势是频繁但价值较低的收购，以及价值明显较高的 IPO 数量较少。“这也与当前宏观环境中的流动性状况有关。”\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">值得一提的是，早在2022年ChatGPT横空出世后，OpenAI在AI技术上突飞猛进，目前ChatGPT周活跃用户达8亿，8个月内增加了一倍。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">近期，OpenAI完成了一笔66亿美元融资，目前估值为5000亿美元，超越马斯克的SpaceX，成为全球最有价值的初创公司。\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">另据The Information报道，\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">今年前七个月，OpenAI收入大约翻了一番，预计年收入将达到120亿美元。同时，OpenAI已签下总额近1万亿美元的算力采购协议，有望成为全球赚钱能力最强的AI公司。\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【新闻来源】钛媒体APP \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.toutiao.com\u002Farticle\u002F7559070602288480783\u002F?upstream_biz=doubao&source=m_redirect\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">https:\u002F\u002Fwww.toutiao.com\u002Farticle\u002F7559070602288480783\u002F?upstream_biz=doubao&source=m_redirect\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（本网转发此文章，旨在为读者提供更多的信息资讯，所涉内容不构成投资、消费建议。文章事实如有疑问，请与有关方核实，文章观点非本网观点，仅供读者参考。）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002F0410478a13374e7cb94549695c203535\u002FAI领域.jpg","https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002Fthumbs\u002F0410478a13374e7cb94549695c203535\u002FAI领域.jpg",0,1,58,"2025-10-11 18:00",2,false,{"id":17,"name":20,"enName":21},"芯位视野","Xinwei Vision","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3A4c2cc533-7c29-4701-9eb3-e94604b933aa%3A0.wav?Expires=1760181562&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=9cZ4rOHLmP5COr7AZeWxVklL4Fk%3D",14883692,"4c2cc533-7c29-4701-9eb3-e94604b933aa","2025-10-11 16:25","Ant, OpenAI, DeepSeek are all in a frenzy! The strongest domestic trillion-parameter flagship model Ling-1T is open-sourced.","\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002Fc3485400866f4a1fa42488e6504d2867\u002Fc6815124670442f89ecf82c1308173f2~tplv-tt-origin-web_gif.jpeg\" width=\"552\" height=\"undefined\" style=\"display: block; margin: auto;\" class=\"ql-align-center\">\u003Cp class=\"ql-align-center\">\u003Cspan style=\"color: rgb(187, 187, 187);\" class=\"ql-lineHeight-1-75\">（Photo credit: taken at the Ant booth at the 2025 World Artificial Intelligence Conference）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px; color: rgb(255, 153, 0);\" class=\"ql-lineHeight-1-75\">Recently, the world's latest trillion-parameter general language model has been open-sourced.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">On October 9th, Ant Group released and open-sourced the trillion-parameter general language model Ling-1T today.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">It is known that Ling-1T is the first flagship model of Ant's BaiLing large model series Ling 2.0, and also the largest and most capable non-thought large model released by Ant BaiLing team so far.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Benchmark test data shows that under limited output token conditions, Ling-1T achieved SOTA performance in multiple complex reasoning benchmarks, and in multiple high-difficulty benchmark tests such as code generation, software development, competition mathematics, professional mathematics, and logical reasoning, Ling-1T achieved leading results, with multiple indicators ranking first among open-source models.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Take the AIME 25 competition math list as an example. Ling-1T achieved an accuracy rate of 70.42% with an average consumption of over 4000 tokens, which is better than Gemini-2.5-Pro, which had an average of over 5000 tokens and an accuracy rate of 70.10%.\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Ant's Ling-1T achieved higher accuracy with fewer tokens, showing an advantage in comprehensive capabilities of reasoning accuracy and thinking efficiency compared to Google's Gemini series.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002Ff7a223e5db994fd0a88e0a14e946a03e\u002Fa6aae0c2d02d4b465162db43fe4fc3cc~tplv-tt-origin-web_gif.jpeg\" width=\"564\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Not only Ant, but before and after the National Day and Mid-Autumn Festival holidays, AI large models from China and the United States, including OpenAI, Alibaba, Zhizhi, and DeepSeek, have accelerated their competition.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">OpenAI released the AI video model Sora2 and announced GPT-5 Pro, ChatGPT framework Apps SDK, etc., at the developer conference, drawing global attention;\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">DeepSeek released DeepSeek-V3.2-Exp, fully compatible with domestic computing cards, and improved training and inference efficiency and API pricing;\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Alibaba Tongyi released the new generation of native multimodal large model Qwen3-Omni, as well as the fully open-source Tongyi DeepResearch model, framework, and solutions;\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cul>\u003Cli class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Zhizhi released the flagship model GLM-4.6, which ranked first globally on Hugging Face Trending, first among open-source models on LMArena, and fourth globally, tying with Qwen Max Preview (closed source) for the first in the country.\u003C\u002Fspan>\u003C\u002Fli>\u003C\u002Ful>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">As Li Kai-fu, founder and CEO of Zero One Wu, said, the base foundation model is a \"arms race\" for companies with billions of dollars.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">NVIDIA CEO Huang Renxun pointed out,\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">AI computing demand has significantly increased over the past six months, and the demand for NVIDIA's new architecture Blackwell chips is \"very very high.\" He believes this marks the beginning of a \"new industrial revolution.\"\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">NVIDIA recently announced plans to invest a total of $100 billion in OpenAI over the next decade to support OpenAI's deployment of NVIDIA systems requiring 10 gigawatts of power, equivalent to 4 million to 5 million GPUs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Sam Altman, CEO of OpenAI, stated on Wednesday,\u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The key to technological breakthroughs lies in smarter models, longer context processing capabilities, and better memory systems.\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">AGI is about discovering new knowledge, and the most important measure of AGI is no longer passing a certain test, but rather the ability of AI to begin to \"discover new knowledge\" and expand the boundaries of human knowledge.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong class=\"ql-lineHeight-1-75\" style=\"font-size: 18px; color: rgb(255, 153, 0);\">The strongest domestic trillion-parameter flagship model has arrived\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In recent years, large language models have developed rapidly, especially the DeepSeek trend, sparking widespread discussions in academia and industry about general artificial intelligence (AGI). While mixed-expert (MoE) models perform well in specific tasks, their training depends on high-performance computing resources, which is costly and limits their application in resource-constrained environments.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">The Ling team at Ant believes that although MoE models require large amounts of high-performance AI chips (such as H100 and H800), and resource supply, low-performance accelerators are more accessible and cost-effective. Therefore, the model needs a technical framework that can switch between heterogeneous computing units and distributed clusters. At the same time, in AI Infra, performance optimization is carried out in terms of compatibility and reliability across clusters and devices. The company's goal is to scale the model without using advanced GPUs.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In March this year, the Ling Team of Ant Group used AI Infra technology to develop two open-source MoE models of the BaiLing series, Ling-Lite and Ling-Plus,\u003C\u002Fstrong>\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">the former has 16.8 billion parameters, while the Plus base model has up to 290 billion parameters, which has attracted widespread attention.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Subsequently, Ant also announced the voice AI framework Ming-UniAudio, the new generation of inference model Ring-flash-2.0, and other model products.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Currently, the AGI team at Ant is mainly led by He Zhengyu, Executive Vice President and CTO of Ant Group. It is reported that He Zhengyu holds a Ph.D. in Computer Science from the Georgia Institute of Technology; he worked at Google from 2012 to 2018, where he founded and led the open-source project gVisor; in 2018, He Zhengyu joined Ant Group, responsible for the company's technical infrastructure, and led important projects such as Ant's cloud-native transformation, green computing practices, confidential computing innovation, and open-source strategic layout. The latest BaiLing large model plan focuses on scenarios such as lifestyle services, financial services, and healthcare.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Now, the BaiLing large model team at Ant has finally announced the latest, globally strong trillion-parameter non-thought model Ling-1T.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Ling-1T uses the Ling 2.0 architecture, pre-trained on 20T+ high-quality, high-reasoning concentration text, supports a maximum context window of 128K, and greatly enhances the model's efficient thinking and accurate reasoning capabilities through an evolutionary chain of thought (Evo-CoT) of \"middle training + post training\".\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Notably, Ling-1T uses FP8 mixed precision training throughout (some technologies have been open-sourced), making it the largest base model known to use FP8 training. This design brings significant memory savings, more flexible parallel partitioning strategies, and a 15%+ end-to-end acceleration.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">During the reinforcement learning phase, the BaiLing team at Ant innovatively proposed the LPO method (Linguistics-Unit Policy Optimization, LingPO), a strategy optimization algorithm with \"sentence\" as the granularity, providing key support for the stable training of the trillion-parameter model. In addition, the BaiLing team proposed a new hybrid reward mechanism, continuously improving the visual aesthetics perception of this trillion-base model while ensuring code correctness and functionality completeness.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">On the ArtifactsBench front-end capability benchmark, Ling-1T scored 59.31 and ranked first among open-source models in the field of visualization and front-end development tasks with a significant advantage.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cimg alt=\"undefined\" src=\"https:\u002F\u002Fimage.51xinwei.com\u002F2025\u002F10\u002F6dcda3f485a34882a0737bafc8172d7f\u002F8fa1faf093335c494e0fe416984784f1~tplv-tt-origin-web_gif.jpeg\" width=\"692\" height=\"undefined\" style=\"display: block; margin: auto;\">\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">According to the BaiLing team at Ant, in addition to the non-thought model Ling-1T, the BaiLing team is also training a trillion-parameter deep-thinking large model called Ring-1T, which has already open-sourced a preview version on September 30th.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Currently, developers can experience the Ling-1T model.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Raised 192.7 billion USD, global AI investment sets record\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">According to PitchBook statistics, this year, global AI startup companies have attracted a record 192.7 billion USD in venture capital, and 2025 is expected to be the first year where more than half of the venture capital funds flow into the AI industry.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Most of the venture capital funding went to relatively mature startups. Among them, top AI startups like OpenAI, Anthropic, and xAI raised hundreds of billions of dollars this quarter, while some less-known new startups struggled, especially those not focused on AI. PitchBook found that the tightening IPO and M&A environment also made some venture capitalists unwilling to bet on unproven companies.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">In the last quarter, US venture capital invested 62.7% of its funds in AI-related companies, while the global figure was 53.2%. So far this year, the total global venture capital has reached 366.8 billion USD, with US venture capital reaching 250.2 billion USD.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Kyle Sanford, research director at PitchBook, said, \"The market is diverging - either doing AI or not, either big companies or not.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Dimitri Zabelin, senior research analyst at PitchBook for AI and cybersecurity, said that the current main exit trend is frequent but low-value acquisitions, and fewer IPOs with significantly higher value. \"This is also related to the liquidity situation in the current macro environment.\"\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Notably, after the emergence of ChatGPT in 2022, OpenAI made rapid progress in AI technology, and currently ChatGPT has 800 million weekly active users, doubling in eight months.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">Recently, OpenAI completed a $6.6 billion financing round, and its valuation now stands at $500 billion, surpassing Elon Musk's SpaceX, becoming the most valuable startup in the world.\u003C\u002Fspan>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cbr>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">According to The Information, \u003C\u002Fspan>\u003Cstrong style=\"font-size: 18px;\" class=\"ql-lineHeight-1-75\">OpenAI's revenue for the first seven months of this year has roughly doubled, and is expected to reach $12 billion annually. Meanwhile, OpenAI has signed nearly $1 trillion in computing power purchase agreements, and is expected to become the most profitable AI company in the world.\u003C\u002Fstrong>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>\u003Cp>\u003Cspan style=\"color: rgb(187, 187, 187);\">【News Source】 Titanium APP \u003C\u002Fspan>\u003Ca href=\"https:\u002F\u002Fwww.toutiao.com\u002Farticle\u002F7559070602288480783\u002F?upstream_biz=doubao&source=m_redirect\" rel=\"noopener noreferrer\" target=\"_blank\" style=\"color: rgb(187, 187, 187);\">https:\u002F\u002Fwww.toutiao.com\u002Farticle\u002F7559070602288480783\u002F?upstream_biz=doubao&source=m_redirect\u003C\u002Fa>\u003C\u002Fp>\u003Cp class=\"ql-align-justify\">\u003Cspan style=\"color: rgb(187, 187, 187);\">（This article is reprinted by this site to provide readers with more information and news. The content does not constitute investment or consumer advice. If there are any questions about the facts of the article, please verify with the relevant parties. The views of the article are not the views of this site, and are for reference only.）\u003C\u002Fspan>\u003C\u002Fp>\u003Cp>\u003Cbr>\u003C\u002Fp>","https:\u002F\u002Fxinwei-dev-test.oss-cn-shenzhen.aliyuncs.com\u002Fintelligent\u002Faudio%3Abeb8bccd-966d-4477-8b31-c59d5e21e2de%3A0.wav?Expires=1774838461&OSSAccessKeyId=LTAI5tNvY2RkKjZw4LLWsrPK&Signature=%2Bwgkg%2BCaoLQMPif8Ph9frDfQIAs%3D","beb8bccd-966d-4477-8b31-c59d5e21e2de",18315040]