文章目录AI科技热点日报 | 2026年8月7日 今日摘要一、大模型军备赛参数规模再上台阶事件概要事件概要来源 / Sources二、AI 商业化兑现阿里 Qwen 分成机制与宇树科技 IPO事件概要事件概要来源 / Sources三、智能体进入 L3 时代阶跃星辰判断与商业化博弈事件概要来源 / Sources四、具身智能从舞台到工厂数据成为生死线事件概要事件概要来源 / Sources五、具身数据生态加速黎曼动力、光轮智能与诺亦腾机器人战略协同事件概要来源 / Sources六、国际大模型前沿OpenAI Astra 最快下周发布事件概要来源 / Sources行业数据速览媒体人解读参考来源英文来源中文来源AI科技热点日报 | 2026年8月7日综合整理自路透社、IT之家、晚点 LatePost、TMTPost、钛媒体、第一财经、财联社、DoNews、21世纪经济报道、每日经济新闻等科技财经媒体当日报道聚焦大模型迭代、商业化兑现、具身智能与机器人产业、芯片与算力、企业战略与融资动态。 今日摘要2026年8月7日国内大模型赛道继续在参数规模与商业化两条线上同步推进路透社披露阿里拟对下一代开源 Qwen 大模型的商用用户收取营收分成开源模型变现路径迎来关键转折据晚点 LatePost 报道字节跳动正讨论训练参数规模超 5 万亿的超大模型将超过阿里 Qwen 3.8-Max 与月之暗面 K3成为目前国内公开参数体量最大的模型智谱新一代 GLM-5.3 大模型临近发布参数规模有望从 GLM-5.2 的 7400 亿级别跃升至万亿以上。国际侧OpenAI 被曝最快下周发布代号 “mewfour” 的 Astra 模型是其自 GPT-4.5 以来训练的最大模型。具身智能方面宇树科技科创板 IPO 发行价定为 150.80 元/股、上市市值约 609.99 亿元成为 A 股人形机器人第一股阶跃星辰联合创始人朱亦博在沙利文峰会上判断行业已进入AI L3 智能体时代百度沈抖指出数据是具身智能的生死线。On August 7, 2026, the domestic large-model race kept pushing forward on both the parameter-scale and commercialization fronts: Reuters reported that Alibaba plans to take a revenue share from commercial users of its next-generation open-source Qwen models, marking a pivotal turn in monetizing open-source AI. According to LatePost, ByteDance is discussing training an ultra-large model with over 5 trillion parameters, surpassing Alibaba’s Qwen 3.8-Max and Moonshot AI’s K3 to become the largest disclosed-parameter model in China. Zhipu’s next-generation GLM-5.3 is nearing release, with parameters expected to leap from GLM-5.2’s 740 billion range to over one trillion. On the international side, OpenAI was reported to launch the “mewfour” Astra model as early as next week, the largest model the company has trained since GPT-4.5. In embodied AI, Unitree Robotics set its STAR Market IPO price at RMB 150.80 per share with a market cap of about RMB 60.999 billion, becoming the first A-share humanoid-robot listing. StepFun co-founder Zhu Yibo judged at the Frost Sullivan summit that the industry has entered the “AI L3 Agent era”, while Baidu’s Shen Dou pointed out that data is the “life-or-death line” of embodied intelligence.一、大模型军备赛参数规模再上台阶事件概要8月7日据《晚点 LatePost》报道字节跳动正讨论训练一个参数规模超 5 万亿的模型体量将超过阿里 Qwen 3.8-Max2.4 万亿参数与月之暗面 K32.8 万亿参数成为目前国内已公开信息中参数量最大的模型。该计划仍处于早期阶段最终是否发布仍存不确定性。新模型将由 Seed Foundation 负责人项亮牵头并联合大语言模型预训练数据负责人沈科共同推进Seed 团队目前正在重新梳理组织、划分职责并分配资源。多位字节人士表示Seed 团队判断与其在现有尺寸上继续追赶同行不如将参数规模一次推到同行的数倍以争取领先位置。On August 7, LatePost reported that ByteDance is discussing training a model with over 5 trillion parameters, exceeding Alibaba’s Qwen 3.8-Max (2.4T) and Moonshot AI’s K3 (2.8T), making it the largest disclosed-parameter model in China. The plan is still at an early stage and final release remains uncertain. The new model will be led by Xiang Liang, head of Seed Foundation, together with Shen Ke, who leads pre-training data work for large language models. ByteDance’s Seed team is currently restructuring its organization and reallocating resources, judging that pushing the parameter scale several times beyond peers in one shot will be more effective than incrementally catching up.事件概要同一天智谱公司创始人确认新一代 GLM-5.3 大模型即将发布市场预期本次升级幅度为史诗级 Plus参数规模有望从 GLM-5.2 的 7400 亿级别跃升至万亿以上性能有望冲击全球前沿水平。GLM-5.2 目前已在 Artificial Analysis 评测中取得 52 分为国产模型中首个在编程能力上比肩美国前沿大模型的产品。机构观点认为大模型竞争已从宏大叙事转向兑现证据头部厂商密集迭代推动 AI 应用商业化进程加速外资亦逆势加仓 AI 硬件板块。On the same day, Zhipu’s founder confirmed that the next-generation GLM-5.3 large model is about to be released, with market expectations of an “epic-plus” upgrade that could push parameters from GLM-5.2’s 740 billion range to over one trillion, with performance aimed at the global frontier. GLM-5.2 currently scores 52 on Artificial Analysis, the first domestic model to match U.S. frontier peers in coding capability. Analysts noted that the large-model race has shifted from “grand narrative” to “evidence of delivery”, with intensive iteration by leading players accelerating AI commercialization, while foreign capital is counter-cyclically increasing positions in AI hardware.来源 / Sources晚点 LatePostDoNews 转载财联社第一财经二、AI 商业化兑现阿里 Qwen 分成机制与宇树科技 IPO事件概要8月7日路透社报道阿里巴巴计划向旗下下一代开源 AI 大模型的大型商用用户收取营收分成并将于下周为自家开源模型推出同款分成机制。在此之前阿里巴巴仅对在自身云计算平台调用模型的开发者收费绝大多数开源模型允许客户部署在自有数据中心、全程免费使用。这一机制如落地将是开源大模型首次系统性引入商用分成模式可能改变当前开源 AI 的商业格局。On August 7, Reuters reported that Alibaba plans to take a revenue share from large commercial users of its next-generation open-source AI models and will roll out a similar mechanism for its open-source models as early as next week. Until now, Alibaba has only charged developers who call its models on Alibaba Cloud itself, while the vast majority of open-source models allowed customers to deploy them on their own data centers for free. If implemented, this would mark the first time an open-source large model systematically introduces a “commercial revenue-share” model, potentially reshaping the economics of open-source AI.事件概要同日人形机器人龙头宇树科技发布科创板上市发行公告确定发行价格为 150.80 元/股发行数量 4044.6434 万股占发行后总股本 10%发行市盈率 219.23 倍高于行业平均 38.56 倍预计募集资金总额约 60.99 亿元、净额约 59.17 亿元对应上市市值约 609.99 亿元。战略配售获配 808.9286 万股包括社保基金、深度求索、中国石油集团等网上申购日为 8 月 10 日缴款日为 8 月 12 日。募集资金主要用于智能机器人模型研发项目20.22 亿元和机器人本体研发项目11.1 亿元。市场普遍认为作为A 股人形机器人第一股宇树科技 IPO 将重塑 A 股机器人板块的估值体系。The same day, humanoid-robot leader Unitree Robotics released its STAR Market IPO prospectus, setting the offering price at RMB 150.80 per share, with 40.4464 million shares issued, accounting for 10% of post-issuance total shares, at a P/E ratio of 219.23 (versus an industry average of 38.56). Total proceeds are expected at approximately RMB 6.099 billion, net RMB 5.917 billion, implying a market cap of about RMB 60.999 billion. Strategic placement of 8.0893 million shares included the National Social Security Fund, DeepSeek and CNPC; the online subscription date is August 10 and the payment date is August 12. Proceeds will mainly fund intelligent robot model RD (RMB 2.022 billion) and robot body RD (RMB 1.11 billion). The market widely views this as the “first A-share humanoid robot listing” and expects it to reset the valuation framework for A-share robotics.来源 / Sources路透社 Reuters界面新闻财联社DoNews三、智能体进入 L3 时代阶跃星辰判断与商业化博弈事件概要8月5日第二十届沙利文全球增长、科创与领导力峰会暨第五届新投资大会在上海举行。阶跃星辰联合创始人、首席技术官朱亦博在演讲中表示行业已经无可置疑地进入了 AI 的 L3 时代。公开资料显示阶跃星辰描绘的 AGI 路线图将能力分为五级L1 聊天机器人、L2 推理者、L3 智能体不仅能思考还可以采取行动的 AI 系统、L4 创新者、L5 组织者。这一时代起点被认为始于去年下半年 Claude 系列模型在长链路任务上的突破。当前全球大模型公司均把 Agent 作为下一代主战场行业普遍认为中美模型差距正在进一步收窄万亿元级资本正围绕真实场景商业化展开博弈。On August 5, at the 20th Frost Sullivan Global Growth, Innovation and Leadership Summit 5th New Investment Conference held in Shanghai, StepFun co-founder and CTO Zhu Yibo stated in his speech: “The industry has unquestionably entered the AI L3 era.” According to publicly available materials, StepFun’s AGI roadmap divides capabilities into five levels: L1 chatbot, L2 reasoner, L3 agent (an AI system that not only thinks but also takes action), L4 innovator, and L5 organizer. The starting point of this era is seen as beginning with Claude’s breakthroughs in long-horizon tasks in the second half of last year. Global large-model companies are now treating agents as the next main battleground, with industry consensus that the U.S.–China model gap is narrowing and trillion-yuan-level capital is now competing around real-scenario commercialization.来源 / Sources每日经济新闻第一财经36氪四、具身智能从舞台到工厂数据成为生死线事件概要8月7日百度智能云事业群总裁沈抖在行业论坛上指出2026 年是人形机器人学会干活的量产元年但当前最大的瓶颈不是硬件而是数据——“数据是制约行业发展的核心短板具身智能尚未进入生产生活环节数据未形成规模化正向循环与自动驾驶的成熟数据生态存在较大差距”。沈抖认为相较于大语言模型具身智能领域的实际投入差了一个数量级产品技术与应用场景拓展都需要加速推进。On August 7, Baidu Intelligent Cloud President Shen Dou pointed out at an industry forum that 2026 is the mass-production year for humanoid robots to “learn to work”, but the biggest bottleneck is not hardware — it is data. “Data is the core shortcoming constraining the industry. Embodied intelligence has not yet entered production and life scenarios, and data has not formed a scaled positive cycle, leaving a large gap compared with the mature data ecosystem of autonomous driving.” Shen Dou noted that compared with large language models, actual investment in embodied intelligence is one order of magnitude lower, and both product technology and scenario expansion need to accelerate.事件概要8月6日合肥市具身智能机器人数据采集预训练场内83 台来自不同企业的机器人在家居、商超、工业等领域的 30 余个复刻场景中开展实训。通过数据采集员一次次手把手教学机器人作业能力不断进化。这一基础设施正是百度沈抖所强调的规模化正向循环路径的落地探索。同日瑞萨电子Renesas宣布将携多款面向具身智能、工业机器人、服务机器人及 AI 开发平台的解决方案首次亮相中国具身智能机器人产业大会暨展览会包括传感器融合灵巧手、机器人关节、9 轴电机控制、永磁同步马达控制解决方案等。On August 6, at the Hefei Embodied Intelligent Robot Data Collection Pre-training Field, 83 robots from different companies were training in more than 30 replicated scenarios covering home, retail, and industrial settings. Through repeated “hand-by-hand” teaching by data collectors, the robots’ operational capabilities continue to evolve. This infrastructure is exactly the on-the-ground exploration of the “scaled positive cycle” emphasized by Baidu’s Shen Dou. The same day, Renesas Electronics announced it will debut solutions for embodied intelligence, industrial robots, service robots, and AI development platforms at the China Embodied Intelligent Robot Industry Conference, including sensor-fused dexterous hands, robot joints, 9-axis motor control, and PMSM control solutions.来源 / Sources腾讯云开发者社区人民财讯财联社DoNews五、具身数据生态加速黎曼动力、光轮智能与诺亦腾机器人战略协同事件概要8月6日黎曼动力宣布与光轮智能、诺亦腾机器人Noitom Robotics达成战略合作围绕 Riemann-1.0 具身世界动作模型与 Matrix-Game 3.5 交互式世界模型三方将在高质量具身数据采集、模型训练、规模化评测、真实机器人部署与反馈优化等方向展开深度协同推进面向 2026 年的百万小时具身智能数据建设。黎曼动力表示将联合各方打通从数据采集、模型训练到真实机器人部署的闭环弥补当前国内具身智能数据生态的结构性缺口。On August 6, Riemann Dynamics announced a strategic partnership with Lightwheel Intelligence and Noitom Robotics. Centered on the Riemann-1.0 embodied world-action model and the Matrix-Game 3.5 interactive world model, the three parties will collaborate deeply on high-quality embodied data collection, model training, scaled evaluation, real-robot deployment and feedback optimization, advancing the construction of one million hours of embodied-intelligence data for 2026. Riemann Dynamics said it will work with the partners to close the loop from data collection and model training to real-robot deployment, addressing the structural gap in China’s embodied-intelligence data ecosystem.来源 / SourcesTechWebDoNews财联社六、国际大模型前沿OpenAI Astra 最快下周发布事件概要8月7日IT 之家援引消息源 synthwavedd 报道OpenAI 目标最快下周发布 Astra 模型内部代号为 mewfour是 OpenAI 自 GPT-4.5 以来训练的最大模型。8月1日 OpenAI 发布的有关数学和理论计算机科学的十项突破即由 Astra 内部版本实现覆盖高维球体堆积密度上限、二进制与球面码界限提升、非纯 sofic 群存在性证明、Connes 刚性猜想的推翻、算术电路复杂性新下界、量子平行重复定理、最近向量问题的多项式因子近似难度证明、Ehrhart 体积猜想、多色 Ramsey 数的超指数下界以及极值数猜想等按照 Sol API 的费率计算找到这些问题的解决方案所需总 token 成本约为 2000 美元。该模型可能在 GPT-5.7、GPT-6 或其他名称下发布分析认为 OpenAI 可能参考 Anthropic 做法——向普通消费者提供一版功能更强大的变体则需经特殊审批才能访问。On August 7, IT Home cited source synthwavedd reporting that OpenAI aims to launch the Astra model as early as next week, internally codenamed “mewfour” — the largest model OpenAI has trained since GPT-4.5. The ten breakthroughs in mathematics and theoretical computer science that OpenAI published on August 1 were achieved by an internal version of Astra, covering upper bounds on high-dimensional sphere packing density, improved binary and spherical-code bounds, a proof of the existence of non-sofic groups, refutation of the Connes rigidity conjecture, new lower bounds for arithmetic circuit complexity, the quantum parallel repetition theorem, a polynomial-factor inapproximability proof for closest vector problem, the Ehrhart volume conjecture, super-exponential lower bounds for multicolor Ramsey numbers, and the extremal number conjecture, etc. At Sol API rates, the tokens spent finding solutions to these problems cost approximately USD 2,000. The model may ship as GPT-5.7, GPT-6, or under another name; analysts suggest OpenAI may follow Anthropic’s approach — providing one version to general consumers while reserving more powerful variants behind special approval.来源 / SourcesIT之家路透社 ReutersThe InformationTechCrunch行业数据速览指标数据备注字节跳动讨论训练模型参数规模超 5 万亿超过 Qwen 3.8-Max2.4T与月之暗面 K32.8T智谱 GLM-5.2 Artificial Analysis 评分52 分国产首个编程能力比肩美国前沿智谱 GLM-5.3 预期参数规模万亿从 GLM-5.2 的 7400 亿级别跃升阿里 Qwen 开源商用分成节奏下周推出开源大模型首次系统性引入分成宇树科技 IPO 发行价150.80 元/股发行市盈率 219.23 倍宇树科技 IPO 募资总额约 60.99 亿元净额约 59.17 亿元宇树科技上市市值约 609.99 亿元A 股人形机器人第一股OpenAI Astra 内部代号mewfour自 GPT-4.5 以来训练的最大模型OpenAI Astra 数学突破算力成本约 2000 美元按 Sol API 费率计算黎曼动力数据建设目标百万小时面向 2026 年具身智能合肥预训练场机器人数83 台30 余个复刻场景实训沙利文峰会 AI 分级L1-L5 五级L3 智能体为当前时代起点媒体人解读8 月 7 日的产业信号非常清晰大模型竞争已从宏大叙事全面切换到兑现证据。在国内端最值得关注的两件事是阿里 Qwen 分成机制与字节 5 万亿参数模型前者是开源大模型商业模式的根本性重构意味着完全免费的开源时代开始让位于分层变现后者则说明头部厂商已经意识到与其在小尺寸上死磕基准跑分不如把参数规模推到同行的数倍以拉开身位——这是 OpenAI/Anthropic 在 GPT-5/Claude Opus 5 阶段已经走过的路径。智谱 GLM-5.3 的史诗级 Plus升级同样是这一逻辑的延伸。在国际端OpenAI Astra 的快速逼近意味着 GPT-4.5 以来最大的预训练成果即将进入产品化阶段。如果按 Sol API 费率仅用 2000 美元就完成了十项数学难题突破那么模型规模 × 推理效率的乘积正在快速收敛模型本身的科研能力权重将持续上升。具身智能赛道则在 8 月 7 日出现了三种代表性信号宇树科技 IPO资本端定价、阶跃星辰 L3 Agent 判断模型端定位、百度沈抖数据生死线基础设施端呼吁。三件事分别对应投融资—模型路线—数据闭环三个相互咬合的环节。A 股人形机器人第一股的诞生意味着板块估值锚定机制正式形成而数据是生死线的判断与黎曼动力-光轮智能-诺亦腾三方合作指向同一结论——未来 12 个月内能跑通数据采集→模型训练→真实部署→反馈回流闭环的厂商将率先跨过具身智能的 L3 门槛。值得提醒读者的是当下行情的最大风险是叙事大于兑现8 月以来概念股上涨节奏明显加快但商业化路径、量产规模、毛利结构仍在验证之中。从大模型到具身智能行业正在从故事驱动过渡到订单驱动。下一阶段建议重点跟踪三条主线①开源大模型的分成与本地化部署平衡点②头部人形机器人厂商的工厂落地数据③国产 AI 芯片在推理侧的实际算力利用率与单位 token 成本曲线。The industry signal on August 7 is crystal clear:the large-model race has fully pivoted from “grand narrative” to “evidence of delivery”. Domestically, the two most notable events are Alibaba’s Qwen revenue-share mechanism and ByteDance’s 5-trillion-parameter model. The former represents a fundamental restructuring of the open-source large-model business model — the era of “completely free” open source is giving way to “tiered monetization”; the latter shows that leading players have realized that, instead of fighting on small-scale benchmarks, pushing the parameter scale several times beyond peers in one shot is the right path to widen the gap — exactly the route OpenAI and Anthropic took at the GPT-5 / Claude Opus 5 stage. Zhipu’s “epic-plus” GLM-5.3 upgrade extends the same logic.Internationally, OpenAI Astra’s rapid approach means that the largest pre-training achievement since GPT-4.5 is about to enter productization. If ten mathematical breakthroughs were completed for only about USD 2,000 in tokens at Sol API rates, then the product of “model scale × inference efficiency” is converging rapidly, and the weight of “scientific capability” within the model itself will continue to rise.In embodied AI, three representative signals emerged on August 7:Unitree’s IPO (capital-side pricing),StepFun’s L3 Agent judgment (model-side positioning), andBaidu Shen Dou’s “data life-or-death line” (infrastructure-side call). The three events correspond to three interlocking links: investment financing, model roadmap, and data loop. The birth of the first A-share humanoid-robot listing means the sector’s valuation anchor mechanism is officially in place. The “data is the life-or-death line” judgment and the Riemann Dynamics – Lightwheel Intelligence – Noitom three-way partnership point to the same conclusion — within the next 12 months, vendors who can run through the “data collection → model training → real deployment → feedback return” loop will be the first to cross the embodied-intelligence L3 threshold.It is worth reminding readers thatthe biggest risk in the current market is “narrative outrunning delivery”: since August the pace of concept-stock rallies has clearly accelerated, but commercialization paths, mass-production scale, and gross-margin structures are still being verified. From large models to embodied intelligence, the industry is transitioning from “story-driven” to “order-driven”. For the next phase, we recommend tracking three main threads: ① the balance point between revenue-sharing and on-premise deployment for open-source large models; ② factory-deployment data from leading humanoid-robot vendors; ③ the actual compute utilization and per-token cost curve of domestic AI chips on the inference side.参考来源英文来源Reuters TechnologyIT 之家 (IT Home)The InformationTechCrunch中文来源晚点 LatePostDoNews 转载财联社第一财经每日经济新闻21世纪经济报道界面新闻36氪钛媒体 TMTPostTechWeb腾讯云开发者社区本文内容综合整理自上述公开信息来源媒体人解读仅代表个人观点仅供参考。作者署名刘一说 ️发布日期2026年8月7日