编辑:前沿在线 编辑部
本周 AI 产业迎来密集爆发期:国产大模型完成新一轮代际升级,阿里 Qwen4正式官宣,美团 1.6 万亿参数模型以 MIT 协议全量开源,百度文心承诺开放权重,字节生图模型向专业情形纵深,模算协同的自主生态持续夯实。
海外巨头同步全速迭代,谷歌一日内连发自研 TPU、机器人模型、旗舰多智能体机型,英伟达升级机器人世界模型与工具体系。资本端动作频频,平头哥拟分拆赴港上市,Grok 进入强化学习冲刺阶段,工艺、开源、产业、资本四线协同推进。
阿里官宣 Qwen4 发布 Qwen4-Max 预览版率先上线
阿里正式官宣新一代千问 Qwen4 大模型系列,Qwen4-Max 预览版同步上线,在通用推理、数学代码、多模态理解、长上下文、智能体执行等核心维度完成代际升级。
作为千问全新旗舰机型线,Qwen4 在多项国际基准实测中追平全球第一梯队,同步优化了端云协同部署效率,后续将逐步释放全系列版本与开源权重。这是国产大模型又一次冲顶之作,依托全栈算力与工程能力,国产模型的代际迭代节奏持续加快。
谷歌一年内第四款 TPU 登场 8tx 主打机密计算 8i Max 双芯亮相
谷歌发布两款新一代自研 TPU 芯片,分别是面向机密计算的 TPU v8tx 与超大规模训练专用的 TPU v8i Max,这已是谷歌一年内推出的第四款 AI 芯片。
v8tx 针对敏感数据推理做硬件级安全隔离,满足金融、政务等情形的合规需要;v8i Max 使用双芯封装架构,算力密度与互联带宽大幅跃升,将率先部署于谷歌云与 Gemini 训练体系。
谷歌正借助高密度的芯片迭代,构建 “芯片 - 云 - 模型” 全栈闭环,对抗英伟达的通用算力生态长处。
英伟达正式发布 GR00T 2 通用机器人世界动作模型,在运动控制、情形泛化、精细操作能力上全面升级,可通配人形、四足、机械臂等多形态机器人本体;同步推出 Isaac Lab 3.0 仿真开发平台,训练数据生成效率提高 3 倍,仿真到真机的策略迁移一致性进一步提高。
GR00T 2 已向核心合作伙伴开放内测,持续完善机器人通用大脑的工艺生态,持续巩固英伟达在具身智能领域的软件底座话语权。
百度文心发布ERNIE 5.5 Flash Preview 11月25日前开源权重
百度正式发布文心 ERNIE 5.5 Flash Preview 轻量旗舰模型,使用稀疏 MoE 架构,在保持接近中阶模型能力的同时,大幅压低推理时延与成本,长上下文、工具调用、多模态能力同步优化。
官方明确承诺 11 月 25 日前开放完整模型权重,使用宽松开源协议,兼容企业与开发者二次定制。国产轻量旗舰阵营再添核心选手,大模型开源普惠的节奏持续加快,将进一步降低企业规模化部署的门槛。
字节发布Seedream 5.5 原生透明图生成2K 编辑单张0.28 元
字节跳动正式推出 Seedream 5.5 图像生成模型,是行业首批原生兼容 Alpha 通道透明背景图生成的机型,同时升级 2K 分辨率图像编辑能力,API 定价单张 2K 编辑低至 0.28 元。
模型在材质还原、细节精度、图文一致性上全面优化,深度适配电商外观、品牌物料、UI 创作等专业生产情形。生图赛道正从通用创作向专业工业情形纵深,高性价比策略持续挤压竞品空间,完善字节 AIGC 内容生态闭环。
谷歌发布 Gemini Robotics 2.5 奔驰产线分拣试点成功率 96%
谷歌发布 Gemini Robotics 2.5 机器人专用模型,强化了视觉感知与动作规划的端到端协同,在奔驰汽车产线的零部件分拣试点中,任务成功率达到 96%,可自适应多品类工件的柔性分拣需要。
模型具备强泛化能力,无需大规模重新训练即可快速迁移至不同工业工位。这标志着通用机器人模型正式切入高端制造核心作业环节,验证了大模型在工业产线的商用落地价值。
阿里拟分拆平头哥赴港独立上市 芯片累计出货超 56 万颗
据市面消息,阿里正推进平头哥半导体分拆独立,筹备赴港上市事宜,当下已启动前期架构调整与辅导工作。平头哥自研 AI 芯片累计出货量已突破 56 万颗,真武系列覆盖云端训练、边缘推理全情形,软硬件生态适配持续完善。
若顺利登陆资本市面,将成为国产 AI 算力赛道又一标杆上市公司。分拆独立是国产 AI 芯片进入资本化兑现期的关键信号,市面化运作将进一步激发业务活力,加速自主算力生态的规模化扩张。
马斯克:Grok 4.8 进入强化学习阶段 本周小范围推送
马斯克透露,xAI 旗下 Grok 4.8 大模型已进入强化学习训练阶段,将于本周启动小范围灰度推送。新版本重点打磨长链逻辑推理与代码工程能力,同时优化实时信息接入的准确性与时效性,后续将逐步向全量订阅使用者开放。
头部大模型的竞速已进入强化学习深水区,借助 RL 阶段持续拉齐能力上限,成为厂商冲刺第一梯队的核心路径。
谷歌旗舰 Gemini 4 Carbon 正式发布 多智能体协作一夜造游戏
谷歌正式发布旗舰模型 Gemini 4 Carbon,主打多智能体协同与长周期复杂工程任务。官方实测表明,由多个 Carbon 模型组成的智能体团队,可在一夜之间自主完成一款完整小游戏从玩法外观、代码开发到美术产出的全流程工作。
模型在网络安全、代码工程、科研辅助等专业情形能力全面升级,此前已登顶多项权威智能实测榜首。多智能体协同完成完整工程项目成为现实,标志着大模型从单点辅助工具升级为全流程生产主体,生产力价值达成阶跃式提高。
美团 LongCat-2.5 正式开源 MIT 协议 1.6 万亿参数全程国产算力训练
美团正式开源 LongCat-2.5 大模型,使用 MIT 宽松开源协议,总参数规模达 1.6 万亿,全程基于国产算力集群完成全流程训练,是当下已开源的最大规模国产算力训练大模型。
模型针对本地生活、即时零售情形深度优化,兼容百万级上下文与多模态理解,开发者可免费商用与二次修改。万亿级大模型开源加纯国产算力训练双重突破,完整验证了国产算力支撑超大规模模型研发、开源的全链路能力,国产 AI 自主生态建设再下一城。
前沿动态前沿大会
前沿人物
本文来源:https://www.163.com/dy/article/LBGEO79505569XIR.html
若有版权问题,请来信告知:13951431913#139.com 或在右方提交评论。
0
Editor: Frontier Online Editorial Department
This week saw a concentrated burst in the AI industry: domestic large models completed another generational upgrade, with Alibaba Qwen4 officially announced, Meituan’s 1.6 trillion parameter model fully open-sourced under the MIT license, Baidu Wenxin committing to open-source model weights, ByteDance’s image-generation model advancing deeper into professional scenarios, and a self-developed ecosystem of model-compute collaboration continuing to be strengthened.
Overseas giants are iterating at full speed in parallel: Google released its in-house TPU, a robotics model, and a flagship multi-agent model within a single day, while Nvidia upgraded its robotics world model and toolset. On the capital front, moves are frequent: T-Head plans a spin-off listing in Hong Kong, Grok has entered the final stretch of reinforcement learning, and process, open-source, industrial, and capital dimensions are advancing in concert across four tracks.
Alibaba announces Qwen4; the Qwen4-Max preview is first released
Alibaba officially announced its new-generation Qwen4 large model series, with the Qwen4-Max preview released at the same time, delivering generational upgrades across core dimensions including general reasoning, math and code, multimodal understanding, long context, and agent execution.
As Qwen’s new flagship model line, Qwen4 matches the global first tier in multiple international benchmark tests, while also optimizing end-cloud collaborative deployment efficiency; a full range of versions and open-source weights will be released gradually. This marks another top-tier push by a domestic large model, and relying on full-stack computing power and engineering capabilities, the pace of generational iteration for domestic models continues to accelerate.
Google’s fourth TPU within a year arrives: 8tx targets confidential computing; 8i Max appears with a dual-chip design
Google released two new-generation in-house TPU chips—TPU v8tx aimed at confidential computing and TPU v8i Max specialized for ultra-large-scale training—making this Google’s fourth AI chip launch within a year.
v8tx provides hardware-level secure isolation for inference on sensitive data, meeting compliance needs in finance, government, and other scenarios; v8i Max adopts a dual-chip packaging architecture, significantly boosting compute density and interconnect bandwidth, and will be deployed first in Google Cloud and the Gemini training system.
Google is leveraging high-density chip iterations to build a full-stack “chips–cloud–models” closed loop, countering Nvidia’s strengths in its general-purpose compute ecosystem.
Nvidia officially released GR00T 2, a general-purpose robotics world action model, delivering comprehensive upgrades in motion control, scenario generalization, and fine manipulation, and supporting multiple robot form factors including humanoid, quadruped, and robotic arms; it also launched the Isaac Lab 3.0 simulation development platform, increasing training data generation efficiency by 3x and further improving sim-to-real policy transfer consistency.
GR00T 2 has opened beta testing to core partners, continuing to refine the process ecosystem around the robotics general brain and further consolidating Nvidia’s influence over the software foundation in embodied AI.
Baidu Wenxin releases ERNIE 5.5 Flash Preview; weights to be open-sourced before November 25
Baidu officially released the Wenxin ERNIE 5.5 Flash Preview lightweight flagship model, using a sparse MoE architecture to keep capabilities close to mid-tier models while sharply reducing inference latency and cost; long context, tool calling, and multimodal capabilities have also been optimized.
Officials have clearly committed to releasing the full model weights before November 25 under a permissive open-source license, enabling secondary customization for enterprises and developers. A core contender has been added to the domestic lightweight flagship lineup, and the pace of open, inclusive large-model development is accelerating, further lowering the barrier for enterprises to scale deployments.
ByteDance releases Seedream 5.5 with native transparent image generation and 2K editing for 0.28 yuan per image
ByteDance has officially launched the Seedream 5.5 image generation model, one of the industry's first models natively compatible with Alpha-channel transparent background image generation. It also upgrades 2K-resolution image editing capabilities, with API pricing for single 2K image edits as low as 0.28 yuan.
The model has been comprehensively optimized in material reproduction, detail precision, and text-to-image consistency, deeply adapting to professional production scenarios such as e-commerce visuals, brand materials, and UI creation. The text-to-image generation track is evolving from general creation toward specialized industrial applications, with its cost-effective strategy continuously squeezing competitors' market space and completing ByteDance's AIGC content ecosystem loop.
Google releases Gemini Robotics 2.5 with a 96% success rate in Mercedes-Benz production line sorting pilot
Google has released Gemini Robotics 2.5, a model dedicated to robotics, which strengthens end-to-end synergy between visual perception and motion planning. In a parts sorting pilot on a Mercedes-Benz automotive production line, the task success rate reached 96%, demonstrating its ability to adapt to flexible sorting demands for multi-category workpieces.
The model possesses strong generalization capabilities and can be rapidly transferred to different industrial workstations without large-scale retraining. This marks the official entry of general-purpose robot models into core operations of high-end manufacturing, validating the commercial value of large models in industrial production lines.
Alibaba plans to spin off T-Head for an independent listing in Hong Kong; cumulative chip shipments exceed 560,000 units
According to market reports, Alibaba is advancing the independent spin-off of T-Head Semiconductor and preparing for a Hong Kong listing, having already initiated preliminary structural adjustments and advisory work. T-Head's self-developed AI chip cumulative shipments have surpassed 560,000 units, with the Zhenwu series covering all scenarios from cloud training to edge inference, and the software and hardware ecosystem adaptation continues to improve.
If it successfully enters the capital markets, it will become another benchmark listed company in the domestic AI computing track. The independent spin-off is a key signal that domestic AI chips are entering a phase of capital realization; market-oriented operations will further stimulate business vitality and accelerate the large-scale expansion of the self-developed computing ecosystem.
Musk: Grok 4.8 enters reinforcement learning phase, small-scale rollout this week
Musk revealed that xAI's Grok 4.8 large model has entered the reinforcement learning training phase and will initiate a small-scale grayscale rollout this week. The new version focuses on refining long-chain logical reasoning and code engineering capabilities while optimizing the accuracy and timeliness of real-time information integration, and will gradually be rolled out to all subscribed users later.
The race among top-tier large models has entered the deep waters of reinforcement learning; leveraging the RL phase to continuously raise the capability ceiling has become the core path for manufacturers to rush to join the top tier.
Google's flagship Gemini 4 Carbon officially released: multi-agent collaboration builds a game overnight
Google has officially released its flagship model Gemini 4 Carbon, highlighting multi-agent collaboration and long-cycle complex engineering tasks. Official testing shows that a team of multiple Carbon model agents can autonomously complete the entire workflow of a complete mini-game—from gameplay design, code development, to art production—overnight.
The model's capabilities have been comprehensively upgraded in professional scenarios such as cybersecurity, code engineering, and scientific research assistance, and it has previously topped multiple authoritative intelligent benchmarks. Multi-agent collaboration to complete full engineering projects has become a reality, marking the upgrade of large models from single-point assistance tools to full-process production entities, bringing a quantum leap in productivity value.
Meituan officially open-sourced LongCat-2.5 under the MIT license; a 1.6 trillion parameter model trained entirely on domestic computing power.
Meituan officially open-sourced the LongCat-2.5 large model under the permissive MIT open-source license. With a total parameter scale reaching 1.6 trillion, the entire process was completed on a domestic computing power cluster. It is currently the largest domestic-computing-power-trained large model that has been open-sourced.
The model is deeply optimized for local services and instant retail scenarios, compatible with million-level contexts and multimodal understanding. Developers are allowed to use it commercially for free and make secondary modifications. The dual breakthrough of open-sourcing a trillion-level large model and training on purely domestic computing power fully validates the all-chain capability of domestic computing power to support the development and open-sourcing of ultra-large-scale models, marking another significant step forward in the construction of an independent domestic AI ecosystem.
Cutting-edge dynamics Cutting-edge conferences Cutting-edge figures
/ Read the next issue /
Return to NetEase Homepage Download the NetEase News App











