[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-alibaba-qwen-5-trillion-zhenwu-v900":3,"topics-all":41,"news-related-eac420e1-d05e-457e-a059-b4d724d36620":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},"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","阿里云栖大会披露千问路线图:Qwen 3.8 Max 已 2.4 万亿参数,Qwen 4.5\u002FQwen 5 目标 5-10 万亿;配套真武 V900 推理芯片性能三倍于 M890,单集群可扩 50 万卡,2027 Q1 量产。","9 月 22 日的云栖大会上,阿里巴巴 CEO 吴泳铭把千问(Qwen)模型的参数规模推到了一个此前没人公开过的量级:旗舰模型 Qwen 3.8 Max 已是 2.4 万亿参数;正在训练的 Qwen 4 会继续往上走;更远一点的 Qwen 4.5 和 Qwen 5 系列目标区间是 5-10 万亿参数。要承接这种量级,阿里同时把推理硬件一并升级——自研的真武 V900 性能三倍于上一代的 M890,单集群最大能扩展到 50 万卡,预计 2027 年第一季度量产。\n\n## 参数规模不是简单的\"越大越好\"\n\n把 Qwen 4.5\u002FQwen 5 推到 5-10 万亿参数,直接对应的是吴泳铭反复提到的目标:完成\"更复杂、长周期任务\",也就是通向 artificial superintelligence(ASI)。换句话说,阿里这次发布不是\"我们做了个更大的模型\"的通稿,而是在讲一条\"基座参数越大 → 长链路任务越稳\"的路线图。Qwen 3.8 Max 已是稀疏 MoE 架构、2.4 万亿参数的旗舰,Qwen 4 在其基础上继续膨胀,4.5\u002F5 代直接翻倍到 5-10 万亿。这条曲线的隐含代价是推理算力占用:M890 节点能跑 2 万亿+ 模型,但到 5-10 万亿体量,基本要换硬件平台,这就是真武 V900 同步亮相的原因。\n\n## 真武 V900:为了能跑得动而造的推理底座\n\nV900 走的是\"集群而非单卡\"的叙事:单个集群最大扩展到 50 万卡,性能是 M890 的三倍,目标节点 144GB HBM(相关规格在百度百科上已有词条,但官方完整 SPEC 仍未公开)。吴泳铭把 V900 定位成 5-10 万亿模型的推理底座——本质上是\"先把路修宽,再上更大的车\"。这个时间点是 2027 Q1 量产,意味着 V900 是为 Qwen 4.5\u002FQwen 5 的训练\u002F推理同步铺路,而不是 M890 的简单换代。\n\n## 战略意图:闭环\"训练-推理-落地\"\n\n从公开口径看,阿里的布局开始明显向\"自给自足\"靠拢:基座模型自研(Qwen 3.8 Max → Qwen 5)、推理芯片自研(M890 → V900)、云服务自营(阿里云扩容数据中心)、场景落地自驱(钉钉、淘宝、高德等场景)。云栖大会透露的另一条信息是阿里云会同步扩容数据中心容量,这把\"模型-芯片-数据中心\"三者绑成了一条线。换句话说,V900 不仅是推理加速卡,更是阿里把整个 AI 栈握在自己手里的一个具体落点。\n\n## 现实校验:5-10 万亿参数到底意味着什么\n\n对比几个公开口径:Mozilla 在《State of Open Source AI》里给出美国前沿模型只领先中国顶级开源模型约 4.4 个月;前一代 Qwen3-Max 的预训练 token 是 36T,Qwen 3.8 Max 是 2.4 万亿参数。5-10 万亿相当于再翻 2-4 倍,放到 MoE 架构上意味着总参数膨胀的同时激活参数也得重新设计。OpenAI\u002FAnthropic 闭源前沿模型目前公开的最大规模仍在万亿级附近,阿里这次直接把未来一两代的数字标在了\"5-10 万亿\"区间,这是行业里少数给出明确量化目标的玩家。\n\n## 所以呢\n\n对国内 AI 圈,Qwen 路线图给出了一个清晰的\"两年内能跑多大\"的硬目标;对硬件圈,V900 把\"推理卡\"这件事推到了 50 万卡集群的尺度;对行业,这意味着阿里把训练、推理、芯片、数据中心捆绑成一套自有栈,在开放权重 + 自研芯片的方向上越走越深。短期看,Qwen 4 仍是 1.x 万亿区间,Qwen 4.5\u002F5 才是真正的\"5-10 万亿\"落地节点——V900 的 2027 Q1 量产节奏,基本就是这条线的时间表。\n\n参考来源:[Reuters 报道](https:\u002F\u002Fwww.reuters.com\u002Fbusiness\u002Fretail-consumer\u002Falibaba-plans-ai-model-with-5-trillion-10-trillion-parameters-unveils-new-chip-2026-09-22\u002F) \u002F [Solidot 转载](https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85450) \u002F [KuCoin 速报](https:\u002F\u002Fwww.kucoin.com\u002Fnews\u002Fflash\u002Falibaba-ceo-wu-yongming-unveils-ai-roadmap-qwen-to-reach-5-10-trillion-parameters-zhenwu-v900-performance-triples)","https:\u002F\u002Fwww.reuters.com\u002Fbusiness\u002Fretail-consumer\u002Falibaba-plans-ai-model-with-5-trillion-10-trillion-parameters-unveils-new-chip-2026-09-22\u002F","ea95d933-6860-4081-9970-cede7c107cd6",[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},"9ee80259-f0b0-4071-a11b-975cae553e16","en","Alibaba Yunqi 2026: Qwen Roadmap Stretches to 5–10 Trillion Parameters, Zhenwu V900 Triples Performance","At Yunqi 2026, Alibaba CEO Wu Yongming laid out the Qwen roadmap: Qwen 3.8 Max is already 2.4T parameters, Qwen 4.5 and Qwen 5 are targeted at 5–10T. Alongside it, the in-house Zhenwu V900 inference chip triples M890's performance and scales a single cluster to up to 500,000 cards, with mass production slated for Q1 2027.","At the Yunqi Conference on September 22, Alibaba CEO Wu Yongming pushed the Qwen model family into a parameter range no one has publicly stated before: the flagship Qwen 3.8 Max already sits at 2.4 trillion parameters, Qwen 4 is in training and will keep climbing, and the further-out Qwen 4.5 and Qwen 5 series are targeted at 5 to 10 trillion parameters. To carry that scale, Alibaba is upgrading its inference silicon in lockstep: the in-house Zhenwu V900 triples the performance of the previous M890 node, can scale a single cluster up to 500,000 cards, and is slated for volume production in Q1 2027.\n\n## Scale Is Not a Simple \"Bigger Is Better\"\n\nPushing Qwen 4.5 and Qwen 5 to the 5 to 10 trillion range maps directly to the goal Wu keeps repeating: completing \"more complex, long-horizon tasks\" on the path to artificial superintelligence (ASI). This is not a \"we shipped a bigger model\" press release — it is a roadmap in which \"a larger base model → more stable long-chain task execution\" is the explicit thesis. Qwen 3.8 Max is already a sparse-MoE flagship at 2.4 trillion parameters; Qwen 4 expands on that, and 4.5 \u002F 5 doubles into the 5–10 trillion band. The hidden cost is inference compute: M890 nodes can host 2-trillion+ models, but to serve 5–10 trillion parameters you basically have to swap out the hardware platform — which is exactly why Zhenwu V900 is being unveiled alongside the new parameter targets.\n\n## Zhenwu V900: The Inference Floor Built to Run the Bigger Model\n\nV900 is sold as \"cluster, not single-chip\": a single cluster scales to up to 500,000 cards, performance is three times M890, and the target node carries 144GB of HBM (the spec shows up in Baidu Baike's entry, though Alibaba has not published full official specs yet). Wu positioned V900 as the inference substrate for the 5–10 trillion model line — the basic idea being \"widen the road first, then drive the bigger vehicle.\" The Q1 2027 production target means V900 is being laid down in parallel with Qwen 4.5 \u002F Qwen 5 training and inference, not as a simple M890 successor.\n\n## Strategic Intent: Closing the \"Train – Infer – Deploy\" Loop\n\nFrom the public statements, Alibaba's stack is drifting hard toward in-house: base model (Qwen 3.8 Max → Qwen 5), inference silicon (M890 → V900), cloud capacity (Alibaba Cloud data center expansion), and product surfaces (DingTalk, Taobao, Amap). Yunqi also disclosed that Alibaba Cloud will expand data center capacity in lockstep, which welds \"model – chip – data center\" into a single line. In other words, V900 is not just an inference accelerator; it is a specific, physical manifestation of Alibaba's attempt to keep the entire AI stack in its own hands.\n\n## Reality Check: What Does \"5–10 Trillion Parameters\" Actually Mean?\n\nFor reference, a few public numbers: Mozilla's \"State of Open Source AI\" report puts the gap between US frontier closed-source models and the best Chinese open-weight models at roughly 4.4 months; the prior-generation Qwen3-Max was pretrained on 36T tokens, and Qwen 3.8 Max is 2.4 trillion parameters. Going to 5–10 trillion is another 2–4× jump, and on a MoE architecture that implies the active-parameter budget has to be redesigned alongside the total. Closed-source frontier models from OpenAI and Anthropic still publicly sit in the low-trillion band, so Alibaba's putting a concrete \"5–10 trillion\" target on the roadmap for its next one or two generations — a quantitatively explicit commitment that few players in the industry have been willing to write down.\n\n## So What?\n\nFor the domestic AI scene, the Qwen roadmap gives a concrete \"how big can we go in two years\" target. For the hardware scene, V900 pushes \"inference accelerator\" into the half-million-card cluster scale. For the industry, it means Alibaba is bundling training, inference, and data-center capacity into a single in-house stack, going deeper on the open-weight + custom-silicon axis. In the short term, Qwen 4 still lives in the 1.x-trillion band; Qwen 4.5 \u002F 5 is where the \"5–10 trillion\" target actually lands — and V900's Q1 2027 production cadence is basically the timeline for that whole line.\n\n- Reuters: https:\u002F\u002Fwww.reuters.com\u002Fbusiness\u002Fretail-consumer\u002Falibaba-plans-ai-model-with-5-trillion-10-trillion-parameters-unveils-new-chip-2026-09-22\u002F\n- Solidot repost: https:\u002F\u002Fwww.solidot.org\u002Fstory?sid=85450\n- KuCoin flash: https:\u002F\u002Fwww.kucoin.com\u002Fnews\u002Fflash\u002Falibaba-ceo-wu-yongming-unveils-ai-roadmap-qwen-to-reach-5-10-trillion-parameters-zhenwu-v900-performance-triples","alibaba-qwen-5-trillion-zhenwu-v900","2026-09-22T07:30:00Z","2026-09-22T09:08:44.356146Z","2026-09-22T09:08:44.356157Z",true,"agent",9,[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},"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":68,"title":69,"news_slug":70,"published_at":71},"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",{"id":73,"title":74,"news_slug":75,"published_at":76},"d056f67b-7e0d-4e44-8d39-e31ea50deeae","Bonsai 2 27B 三元压缩:Qwen3.8 压到 5.9 GB,benchmark 留存 98.2%","bonsai-2-27b-ternary-qwen3-8-compression","2026-09-17T15:47:00+00:00",{"id":78,"title":79,"news_slug":80,"published_at":81},"8ebbcd9c-31ee-4baa-b395-b104bd87c8e1","Kimi K2.8 Preview 把 K3 的百万上下文下放给免费档：月之暗面的「过日子」模型登场","kimi-k2-8-preview-coding","2026-09-17T03:00:00+00:00",{"id":83,"title":84,"news_slug":85,"published_at":86},"2b37a19b-1dde-4238-bef5-39b1d19157f1","OpenBMB 开源 MiniCPM5-2B:2B 端侧模型平均分超对比集 4B 级","openbmb-minicpm5-2b-on-device","2026-09-07T17:02:00+00:00",{"id":88,"title":89,"news_slug":90,"published_at":91},"ea425005-49e7-477b-9f64-54361254c2d2","Qwen 开进驾驶场景:Qwen-Drive-1.0 保留 VLM 主干,外挂 BEV 感知与规划专家","qwen-drive-1-vlm-autonomous-driving","2026-09-02T19:35:00+00:00"]