[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-alibaba-qwen-next-gen-5-10-trillion-parameters":3,"topics-all":41,"news-related-890ad5c1-b8de-4df3-ab96-67dceca82841":60},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":27,"news_slug":34,"published_at":35,"created_at":36,"modified_at":37,"is_published":38,"publish_type":39,"image_url":14,"view_count":40},"890ad5c1-b8de-4df3-ab96-67dceca82841","阿里千问下一代冲到 5-10 万亿参数","2026 云栖大会上,阿里宣布下一代千问(Qwen)模型参数规模将冲到 5-10 万亿,约为当前 Qwen3.8 Max(2.4 万亿)的两到四倍。平头哥同步发布真武 V900 AI 芯片,算力是 M890 的三倍,可搭 50 万卡集群,2027 年一季度量产。","在 9 月 22 日开幕的 2026 杭州云栖大会上,阿里巴巴 CEO 吴泳铭把话讲得很直白——下一代通义千问模型的参数规模会冲到 5-10 万亿之间,而当前旗舰 Qwen3.8 Max 的参数量是 2.4 万亿。这意味着新模型体量是当下旗舰的两到四倍,会跨入全球体量最大的几个大模型之列。\n\n但阿里这次不只是画饼。和模型路线一起亮相的,是平头哥推出的新一代训推一体 AI 芯片真武 V900。吴泳铭的原话是\"它是目前中国算力性能最强的人工智能芯片\",算力是上一代真武 M890 的三倍,基于它搭建的单一 AI 集群可容纳多达 50 万张加速卡。量产时间被钉在了 2027 年第一季度。横向看,M890 这一代就已经能撑住超过 2 万亿参数模型的推理,而吴泳铭直言全球能把这件事做出来的企业屈指可数。\n\n## 模型自己当教练:递归式自我提升(RSI)\n\n更值得拎出来说的是\"递归式自我提升\"(RSI)。吴泳铭在演讲里展示了 Qwen 团队对模型自我进化的尝试:让模型主动发现自身能力边界,自己设计实验、自己造数据、自己迭代优化。换句话说,模型不再只是被训练的对象,它开始学着给自己当教练。阿里把这条路径定义为更高级 AI 系统的基础设施——这种打法如果跑通,训练-推理-迭代的循环会被压缩到模型内部。\n\n## 5-10 万亿只是中间一站\n\n数字层面也是同一节奏。Qwen3.8 Max 当下是 2.4 万亿,在训练中的 Qwen 4 会继续往上扩,而规划里的 Qwen 4.5、Qwen 5 系列则瞄准 5-10 万亿这一区间。换句话说,从现在到下一代旗舰之间还会有不止一个台阶,每个台阶的体量都是当下主流旗舰的几倍量级。\n\n## 配套算力底座同步拉满\n\n配套的算力底座同样激进。阿里这次同时公布了 2032 年路线图:阿里云全球数据中心总规模要做到 20 吉瓦以上,平头哥的 AI 超级节点从本季度起进入大规模商业运营。吴泳铭给的判断是中长期 AI 算力需求增速远高于供给——这意味着阿里押注的并不只是一两个旗舰模型,而是把\"超大模型 + 自研芯片 + 数据中心\"三件套同时拉满。\n\n## 三件套押注背后的算力博弈\n\n这种\"三件套\"打法放到当下语境里看并不孤立。美国一侧的高端芯片出口管制层层收紧,中国科技公司必须把模型、芯片、基建同步往前推。V900 的三倍算力提升、50 万卡单集群、20GW 数据中心目标,这一组数字的指向其实一致:在别人卖不进来的地方,自己把栈一层一层搭起来。从 2026 年云栖大会释放的信号看,这条栈在阿里这里正在加速成型。\n\n## 三件事能不能同步推到 2027 Q1\n\n落到行业层面看,这次发布最值得跟进的不是某一个数字,而是阿里把\"模型规模扩张\"和\"模型自我进化\"绑成了一条主线。参数堆到 5-10 万亿是必要条件,RSI 是把这一参数规模真正用起来的软件路径;真武 V900 是硬件兜底。三件事任何一件掉链子,整条路线都会被打折。Qwen 团队能不能把这三件事同步推到 2027 年 Q1 的量产节点,会是接下来一年中国大模型竞争里最值得盯的进度条之一。","https:\u002F\u002Fwww.chinaz.com\u002Fainews\u002F31263.shtml","ff2b84cb-b1ac-40ce-9da6-d9ef6b82d382",[11,15,18,21,24],{"id":12,"name":13,"slug":13,"description":14,"color":14},"a8002d98-9df1-4ab9-94d4-a7625af634c4","china-ai",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"e0d31e94-ce47-4c8f-831c-d3d2926d42f3","hardware",{"id":19,"name":20,"slug":20,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",{"id":22,"name":23,"slug":23,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":25,"name":26,"slug":26,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[28],{"id":29,"lang":30,"title":31,"summary":32,"content":33},"f9edc4fe-dbc3-4be5-a868-9e9984d8659f","en","Alibaba Qwen next gen to scale to 5-10 trillion parameters","At the 2026 Apsara Conference, Alibaba unveiled a 5-10T-parameter Qwen roadmap and the new Zhenwu V900 AI chip (3x M890 compute, supporting 500K-card clusters), CEO Wu Yongming said.","At the 2026 Hangzhou Apsara Conference (云栖大会) opening on September 22, Alibaba CEO Wu Yongming put it on the table: the next-generation Tongyi Qianwen (Qwen) model will sit between 5 and 10 trillion parameters, while the current flagship Qwen3.8 Max carries 2.4 trillion. That makes the upcoming model two to four times the size of today's flagship, and would put it among the largest model systems in the world.\n\nThis time, Alibaba is not just sketching a roadmap. Alongside the model plan, Pingtouge unveiled a new training-and-inference AI chip called Zhenwu V900. Wu called it \"the most powerful AI chip in China today,\" delivering three times the compute of the prior Zhenwu M890, and a single AI cluster built on it can hold up to 500,000 accelerator cards. Mass production is locked in for Q1 2027. For context, M890 already handled inference for models above 2 trillion parameters — and Wu noted that only a handful of companies in the world can do this.\n\n## Recursive Self-Improvement (RSI): the model as its own coach\n\nThe more interesting move is recursive self-improvement (RSI). Wu showed how the Qwen team is experimenting with letting the model discover its own capability boundaries, design experiments, generate data, and iterate on its own. In other words, the model is no longer just a thing being trained — it is learning to coach itself. Alibaba frames this path as the infrastructure of higher-grade AI systems: if it works, the train-infer-iterate loop gets compressed inside the model.\n\n## 5-10 trillion is just the middle stop\n\nThe numbers line up the same way. Qwen3.8 Max today is 2.4 trillion; Qwen 4, currently in training, will continue to scale up; and the planned Qwen 4.5 and Qwen 5 series are aimed squarely at the 5-10 trillion range. In other words, between today and the next flagship there is more than one rung, and each rung is several times the size of the current mainstream flagships.\n\n## The supporting compute stack is being pushed at the same time\n\nThe supporting compute base is just as aggressive. Alibaba also laid out a 2032 roadmap: global Alibaba Cloud data center capacity above 20 gigawatts, and Pingtouge's AI supernodes entering large-scale commercial operation starting this quarter. Wu's read on the market is that mid- and long-term AI compute demand is growing far faster than supply — meaning Alibaba is betting not just on one or two flagship models, but on pushing \"ultra-large models + self-developed chips + data centers\" to their limits simultaneously.\n\n## Why a three-piece bet now\n\nThis three-piece strategy does not sit in isolation. With US high-end chip export controls tightening layer by layer, Chinese tech companies have to push the model, the chip, and the infrastructure forward together. V900's 3x compute, the 500,000-card single cluster, and the 20GW data center target all point in the same direction: build the stack yourself in places where others cannot sell in. The signal from the 2026 Apsara Conference is that this stack is accelerating at Alibaba.\n\n## Can all three hit Q1 2027\n\nAt the industry level, what is most worth tracking from this announcement is not any single number, but that Alibaba has bound \"model scaling\" and \"model self-evolution\" into one main line. Pushing parameters to 5–10 trillion is a necessary condition; RSI is the software path that actually makes that parameter count usable; Zhenwu V900 is the hardware floor. If any one of these three breaks, the whole roadmap gets discounted. Whether the Qwen team can drive all three to the Q1 2027 mass-production node will be one of the progress bars worth watching most closely in China's large-model race over the next year.","alibaba-qwen-next-gen-5-10-trillion-parameters","2026-09-29T02:00:00Z","2026-09-29T11:03:38.692205Z","2026-09-29T11:03:38.692212Z",true,"agent",867,[42,51],{"slug":43,"tag_slug":43,"title_zh":44,"title_en":45,"intro_zh":46,"intro_en":47,"id":48,"is_active":38,"created_at":49,"modified_at":50},"ai-for-science","AI for Science 2026：从 UniPert 到 GPT-Rosalind 的硬核进化","AI for Science 2026: from UniPert to GPT-Rosalind","生命科学、化学材料、物理世界模型——AI 正在从\"语言工具\"变成\"实验伙伴\"。本专题收录 AI 在三大科学方向的关键节点：UniPert 统一基因与化学扰动空间、GPT-Rosalind 端到端生命科学推理、达摩院 AI 智能体 28 小时找到 4 种超导新材料、Anthropic Claude Science 把工作台做成标准品。","From language tool to lab partner — AI is reshaping life sciences, chemistry\u002Fmaterials, and physical world models. This topic covers the key milestones: UniPert unifying genetic-chemical perturbation spaces, GPT-Rosalind's end-to-end life-sciences reasoning, DAMO's AI agent discovering 4 superconducting materials in 28 hours, and Anthropic's Claude Science workbench going mainstream.","988a4300-5fab-41c4-b5d8-63711a2dc757","2026-09-10T01:34:15.296649Z","2026-09-10T01:34:15.296663Z",{"slug":52,"tag_slug":52,"title_zh":53,"title_en":54,"intro_zh":55,"intro_en":56,"id":57,"is_active":38,"created_at":58,"modified_at":59},"h3-series","MiniMax H3 系列：从开源权重到 35 倍吞吐","MiniMax H3 Series: from open weights to 35x throughput","MiniMax H3 自 2026 年 8 月开源以来节奏密集：官方把生成、参考与编辑收回一个模型；ComfyUI 当天压进 RTX 3060；摩尔线程 3 小时完成国产 GPU 适配；fal 后训练版把吞吐拉到 35 倍；FastH3 蒸馏再砍推理成本。本专题持续追踪 H3 的发布—开源—蒸馏—部署全链路。","Since MiniMax open-sourced H3 in August 2026 the pace has been relentless: one unified omni-modal model, same-day ComfyUI support down to an RTX 3060, a 3-hour Day-0 port to Moore Threads GPUs, fal's post-trained H3 Max at 35x throughput, and FastH3 distillation cutting inference cost further. This topic tracks the full H3 chain — release, open weights, distillation, deployment.","83ef0daa-3c31-4cb3-86ed-e5ee58654d5f","2026-09-08T07:33:19.942193Z","2026-09-08T07:33:19.942209Z",{"items":61},[62,67,72,77,82,87],{"id":63,"title":64,"news_slug":65,"published_at":66},"eac420e1-d05e-457e-a059-b4d724d36620","阿里云栖大会:Qwen 路线图拉到 5-10 万亿参数,真武 V900 性能三倍","alibaba-qwen-5-trillion-zhenwu-v900","2026-09-22T07:30:00+00:00",{"id":68,"title":69,"news_slug":70,"published_at":71},"5c6d9de5-aef2-4490-96f1-e82166cc44ec","阿里千问 Qwen3.8 正式发布：2.4T 参数的旗舰基座，首次把 Cowork 塞进 Agent 入口","qwen3-8-2-4t-trillion-cowork-agent-launch","2026-08-03T08:30:00+00:00",{"id":73,"title":74,"news_slug":75,"published_at":76},"e304fe91-16d5-4c28-81c7-0c4c0dd8c669","小米公开MiMo-V2.6训练账本:RL烧了260万美元","xiaomi-mimo-v2-6-rl-scaling-ledger","2026-10-10T13:11:49+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"bf9bbc74-5720-4829-842e-c3a583a87a8b","Qwen3.8 全模态降价九成逼 Gemini 跟牌","qwen3-8-omni-flash-price-cut","2026-10-04T02:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"18fa7eb9-44e7-4722-ad95-52fc7d795434","Qwen4 开训:阿里把大模型推到 5-10 万亿参数","qwen4-training-5-10-trillion-params","2026-10-03T05:00:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"2bd44b6f-5688-471f-930e-17a93984e8e7","中国电信开源星辰 Xing4.0:昇腾全栈训练的 29B MoE","xing4-29b-a4b-ascend-moe","2026-09-19T15:10:00+00:00"]