[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-qwen4-training-5-10-trillion-params":3,"topics-all":38,"news-related-18fa7eb9-44e7-4722-ad95-52fc7d795434":57},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"18fa7eb9-44e7-4722-ad95-52fc7d795434","Qwen4 开训:阿里把大模型推到 5-10 万亿参数","9 月 22 日云栖大会,阿里确认 Qwen4 已开训,参数路线扩展至 5–10 万亿,目标指向 ASI。Qwen3.8-Max 在智能体与前端编程双榜拿下第一。","### 从 2.4 万亿到 10 万亿:参数路线一口气拉到极限\n\nQwen3.8-Max 已经在 2.4 万亿参数规模上跑通,这次云栖大会明确:下一代 Qwen4 已经在训练,后续的 Qwen4.5、Qwen5 系列计划把参数扩展到 5–10 万亿区间,直接对标 ASI(artificial superintelligence)。负责人在现场把\"参数扩展仍是通向通用人工智能的核心路径\"重新摆回桌面,这相当于在 Scaling Law 被反复唱衰的当口,给了一记反向投票。\n\n### RSI 让模型自己\"打工\"\n\n比起单纯的参数膨胀,更值得关注的是 Qwen3.8-Max 已经实现的递归式自我改进(Recursive Self-Improvement, RSI):在完全没有人类参与的情况下,系统连续迭代 33 轮,自己搭训练流程、生成训练数据、设计实验、修复缺陷,把 Artificial Analysis 智能体评测得分又拉高了 12.5%。这条链路的意义不在数字,而在\"模型能修改自己的训练流程\"这件事第一次有了工业级证据。配合平头哥真武 V900 芯片\"性能三倍于上一代\"、可扩展至 50 万卡集群、2027 年一季度量产的节奏,RSI 不再是孤立的算法演示,而是模型—芯片—数据中心三层飞轮的具体粘合剂。\n\n### 多模态矩阵全线加压,11 月是关键节点\n\n云栖披露的多模态路线比模型本体更密集:Qwen3.8-Omni-Flash 统一处理视频、音频、图像与文本;Wan3.0 把单次生成视频长度推到 30 秒并支持结构化指令;Qwen-Audio-3.1 覆盖 ASR\u002FTTS\u002F实时翻译,实时翻译字均延迟压到 2.3 秒;Qwen-Image-3.1、HappyOyster-2.0-Preview 等专项模型同步推进。下一代视频生成模型计划 11 月发布,这是 Qwen 第一次在视频生成上明确节点,也是其对标 Veo\u002FSora 的入场券。\n\n### 开源与商业化同时加速\n\nQwen3.8 系列模型发布一个月内下载量突破 5600 万、衍生模型 1900 余个;阿里累计开源 460 余个千问模型、总下载超 30 亿、衍生模型突破 30 万。商业侧,Qwen3.8-Flash 通过注意力机制优化将训练成本降低近 90%,Perplexity、Airbnb、Pinterest、路透社已经基于 Qwen 构建智能代理与自研模型。另一面,阿里 2026 年第二季度 AI 云及算力服务收入约 484 亿元、同比增长 45%,资本开支 677 亿元、同比增长 75%,CEO 吴泳铭给的目标是到 2032 年阿里云运营的全球数据中心规模超过 20 GW。\n\n### 所以呢\n\n参数扩张、RSI、自研芯片、20GW 数据中心,这四件事被同一场发布会串成一条线,已经不能拆开看。Scaling Law 不是被否定了,而是被替换成了\"模型自迭代 + 硬件协同设计 + 算力规模化\"三位一体的新范式。对国内同行来说,Qwen4 开训意味着下一轮旗舰模型的真正起跑线被拉到了 5–10 万亿级,差距不再由模型大小决定,而由\"谁先把 RSI 跑成工业流水线、谁先把芯片-数据中心-模型协同跑通\"决定。","https:\u002F\u002Fwww.itbear.com.cn\u002Fhtml\u002F2026-09\u002F1567656.html","c36a21ac-2a77-421b-9519-1e150695732a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"01598627-1ea6-4b27-a5d8-874971571a71","llm",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",{"id":19,"name":20,"slug":20,"description":14,"color":14},"d11f0044-8aef-487c-bebe-89ce4683a4a3","moe",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"ba99686f-88a3-447e-8d17-89b74df442e6","en","Qwen4 Enters Training as Alibaba Pushes LLMs to 5–10 Trillion Parameters","At its 2026 Apsara Conference, Alibaba confirmed Qwen4 has entered training, with Qwen4.5 and Qwen5 planned to extend parameters into the 5–10 trillion range, targeting ASI. Qwen3.8-Max took first place on both the agent and front-end programming leaderboards.","Alibaba used its 2026 Apsara Conference on Monday to redraw the LLM scaling map. The current flagship Qwen3.8-Max already runs at 2.4 trillion parameters; the newly announced Qwen4 is now in training, while Qwen4.5 and Qwen5 are planned to push the parameter count into the 5–10 trillion range, aimed directly at artificial superintelligence (ASI). The team put the \"parameter scaling is still the core path to AGI\" claim back on the table at a moment when Scaling Law has been repeatedly questioned, effectively a counter-vote.\n\nWhat deserves more attention than raw scale is the recursive self-improvement (RSI) Qwen3.8-Max has demonstrated. With zero human involvement, the system iterated 33 rounds in a row — building its own training pipelines, generating training data, designing experiments and fixing defects — and lifted its Artificial Analysis agent benchmark score by 12.5%. The significance is not the number itself, but that \"a model editing its own training pipeline\" now has industry-grade evidence. Combined with the Pingtouge Zhenwu V900 chip — three times the performance of the previous generation, scalable to clusters of 500,000 cards, in volume from Q1 2027 — RSI is no longer an isolated algorithmic demo but the concrete glue for the model–chip–datacenter three-layer flywheel.\n\nThe multimodal lineup revealed at Apsara is denser than the LLM roadmap itself. Qwen3.8-Omni-Flash unifies video, audio, image and text; Wan3.0 pushes single-shot video generation to 30 seconds and accepts structured instructions; Qwen-Audio-3.1 covers ASR, TTS and real-time translation with per-character real-time translation latency compressed to 2.3 seconds; Qwen-Image-3.1, HappyOyster-2.0-Preview and other specialised models are advancing in parallel. A next-generation video generation model is scheduled for release in November — the first time Qwen has put a clear milestone on video, and its entry pass against the Veo\u002FSora tier.\n\n### Open source and commercialisation accelerating at the same time\n\nWithin a month of launch, the Qwen3.8 series surpassed 56 million downloads with more than 1,900 derivative models. Alibaba has now open-sourced over 460 Qwen models, with cumulative downloads above 3 billion and more than 300,000 derivative models. On the commercial side, Qwen3.8-Flash cuts training cost by nearly 90% through attention-mechanism optimisation; Perplexity, Airbnb, Pinterest and Reuters have built agents and proprietary models on top of Qwen. In Q2 2026 Alibaba's AI cloud and compute-services revenue reached roughly 48.4 billion yuan, up 45% year-on-year, while capex hit 67.7 billion yuan, up 75% year-on-year. CEO Eddie Wu set the target: by 2032, the global datacenter footprint operated by Alibaba Cloud will exceed 20 GW.\n\n### So what\n\nParameter expansion, RSI, in-house silicon and the 20 GW datacenter target are no longer four separate items — they have been woven into a single line by the same announcement. Scaling Law has not been invalidated; it has been replaced by a three-in-one paradigm: model self-iteration + co-designed hardware + compute at utility scale. For domestic peers, Qwen4 entering training means the real starting line for the next flagship cycle has been pulled to the 5–10 trillion tier. The gap will no longer be decided by model size, but by who first turns RSI into an industrial pipeline and who first makes the chip–datacenter–model loop work end to end.","qwen4-training-5-10-trillion-params","2026-10-03T05:00:00Z","2026-10-03T09:03:18.226133Z","2026-10-03T09:03:18.226165Z",true,"agent",721,[39,48],{"slug":40,"tag_slug":40,"title_zh":41,"title_en":42,"intro_zh":43,"intro_en":44,"id":45,"is_active":35,"created_at":46,"modified_at":47},"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":49,"tag_slug":49,"title_zh":50,"title_en":51,"intro_zh":52,"intro_en":53,"id":54,"is_active":35,"created_at":55,"modified_at":56},"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":58},[59,64,69,74,79,84],{"id":60,"title":61,"news_slug":62,"published_at":63},"d941056b-c2e7-42e5-965a-a982c20b1169","Qwen3.8-Flash-Next 架构细节:Gated Residual 多分支残差 + QSA micro-block 稀疏注意力","qwen3-8-flash-next-cost-efficiency-architecture","2026-09-02T02:00:00+00:00",{"id":65,"title":66,"news_slug":67,"published_at":68},"f3c43720-650b-4054-9349-a1386e06c8fc","Le Chonk 把法国拉回非美\u002F美头部:38 分的 Mistral Large 4","mistral-large-4-le-chonk-intelligence-index-38","2026-10-08T03:30:00+00:00",{"id":70,"title":71,"news_slug":72,"published_at":73},"5ddba2a9-781e-4763-b1e0-1e20c6480391","Mistral Large 4:1万亿参数MoE,月底开源","mistral-large-4-1t-moe","2026-10-06T23:10:00+00:00",{"id":75,"title":76,"news_slug":77,"published_at":78},"9fa3427c-cf69-46c7-9720-cd3b646a155b","B站开源35B翻译模型:3B激活,150种语言","bilibili-index-translate-35b-moe","2026-10-04T13:30:00+00:00",{"id":80,"title":81,"news_slug":82,"published_at":83},"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":85,"title":86,"news_slug":87,"published_at":88},"0372db20-feaf-42b8-98bd-e42d9c550306","德国Kolibri开源:78B参数只激活3.46B","aleph-alpha-kolibri-1-open-moe","2026-10-03T19:14:02+00:00"]