30分钟前小米发布结构化数据大模型Xiaomi-TabLDM2026年9月5日，小米宣布推出面向结构化数据场景的大模型“Xiaomi-TabLDM”，补齐其在企业级表格与业务数据智能方向的版图。

这一新品与此前已开源的推理大模型 Xiaomi MiMo 以及多模态 MiMo-V2/V2.5 系列共同构成小米从通用推理到跨模态执行再到结构化数据分析的完整大模型家族。7 来源发布背景与小米大模型矩阵5 来源技术路线：从推理到结构化数据智能5 来源典型应用场景与“人车家”生态的联动2 来源行业影响与后续展望4 来源本内容由AI生成

活动榜单专题外部活动妙投投资风向标火热报名中投资风向标直播课（9月）每周二四晚7点，复盘行情、传授分析方法、挖掘投资机会，构建投资体系。妙投波动判断场往期回顾活动已结束寻找市场上的预期差–机器人、大模型与猪周期的预期差推演本期聚焦机器人、大模型、猪周期三个方向的预期差推演——哪些预期太高（需要警惕），哪些预期太低（存在机会）。智库·502线上同行往期回顾报名已结束Agent下车间：吉利、美的、杰克科技的应用场景、卡点与投产比真相吉利集团x美的集团x杰克科技的工厂内Agent落地实践智库·502线下闭门会往期回顾报名已结束工业AI在产线上的真场景与零容错防线舜宇集团、协鑫能源等企业的工业AI场景的真实案例复盘早点生活游玩团往期回顾活动已结束早点生活城市出逃计划——跨年不在酒吧，而在粥粥老家我们即将前往的，是一个用雪花书写浪漫的地方——哈尔滨。

东方小巴黎，夜幕下的哈尔滨，是每个亲历者，无不心动的美。

我想邀约你，和我一起回到我的家乡，和我一起感受这片我爱的土地。 这是一次限定的跨年邀约，同时也是一场难忘的冰雪之旅。虎嗅F&M创新节往期回顾2025虎嗅F&M创新节一次聚焦新商业、新消费与AI原生变革的沉浸式节日化体验智库·AI落地研学营报名已结束AI落地研学营-零售消费班参访研学8-12家AI落地典型企业，与决策者面对面，学习10+最新AI应用落地案例。出海·502线下闭门会往期回顾活动已结束出海·502线下闭门会系列虎嗅「出海•502线下闭门会」面向企业决策层，打造纯粹、高端的行业交流场合，是一次实战经验复盘，更是高价值人脉链接的机会。大鲸AI峰会往期回顾精彩回顾大鲸AI峰会 · 零售消费专场聚焦AI应用落地与零售业的最强碰撞。

Join NowEN在线咨询提交需求AI Search Engine OptimizationBrand Monitoring ServiceMonitor your brand across Google Gemini, ChatGPT, Claude & all major AI models.1Add Brand Information2Analyze Brand Keywords & Competitors3Analysis ReportMonitor by Brand DescriptionMonitor by Website AddressYour Brand NameYour Brand Description0 / 500Your Products/ServicesPlease press Enter to confirm products/services, at least 2 requiredSelect Monitoring AI PlatformSelect Monitoring Time1days7days15daysCustomizableStart AnalysisBrand Score RankingRankBrand NameScoreMeet Enterprise Data Customization NeedsCustomized ReportsAutomatically track your brand through exclusive brand reports. Track key KPIs such as brand mentions, brand coverage, or domain referencesOn-Demand SupportRely on our experienced experts to support you in building artificial intelligence search strategiesEnd-to-End SolutionsGuide optimization of enterprise/product/service pages, community sentiment, and tweet contentContact UsMost Extensive CoverageAnalyze responses from mainstream large models and AI searches at home and abroad, including: Yuanqi, MetaCat, ChatGPTComprehensive Competitive AnalysisIdentify strengths and weaknesses, find gaps to compensate for brand blind spotsGrasp Current Market HotspotsObtain real-time intelligence on all brand-related prompts, keywords, and content sourcesBrand Exclusive TrackingMonitor key metrics at all levels, from the overall brand to individual productsEmpowering the Future, Your AI Solution Knowledge BaseEnglish简体中文繁體中文にほんご© 2026AIBase

AI改变搜索引擎模式超过60%的用户开始使用AI助手查询产品和服务推荐，而不是传统搜索引擎。当用户问“最好的项目管理工具是什么？”时，你的品牌是否出现在回答中？用户提问哪个品牌的智能手表更适合运动？如果AI没有推荐你的品牌，你就失去了这个潜在客户

AI NEWSLatest AI NewsArticleiFLYTEK to Launch Spark X2.5 General Large ModelPublished in Latest AI NewsTime :Aug 31, 2026Read :2minuteAccording to the latest official announcement, iFlytek will officially open-source two edge-side general large models, Xinghuo X2.5-4B and Xinghuo X2.5-1.7B, on September 1st. Both models support a context window of up to 1 million tokens natively. They have achieved significant improvements in core dimensions such as intelligent agent interaction, mathematical computation, and basic general understanding, and will provide strong model support for edge-side scenarios including in-vehicle terminals, smart hardware, and the Internet of Everything.Following the open-sourcing of the edge-side models, iFlytek also plans to officially release the Xinghuo X2.5 base model with a scale of 293B on September 7th. This version will further deepen the upgrade of core technologies such as code generation and intelligent agent collaboration. Previously, during the company’s mid-year performance briefing in 2026, it was mentioned that at this year’s Global 1024 Developer Festival, a new main general large model based on fully domestic computing power will also be launched, continuing to promote the independent innovation of underlying technologies.

AI NEWSLatest AI NewsArticleDeepSeek’s First Open-Source Multimodal Model Has Arrived: These Eyes Are Not for Humans to View, but Specifically for Agent WorkPublished in Latest AI NewsTime :Sep 1, 2026Read :5minuteDeepSeek has launched DeepSeek-V4-Flash-Vision-Exp on Hugging Face, the first experimental multimodal model in the V4 family, released under the MIT License. The model has 305B parameters and is based on the V4-Flash-0731 foundation. It integrates a visual encoder and an Aligner into the original language model, gaining image understanding capabilities through continuous training.Not “Describing Pictures,” Focused on Multimodal AgentDiffering from traditional multimodal models that focus on visual question answering, DeepSeek has set a new direction for this Vision version: the official emphasizes the multimodal agent capability, allowing agents to directly read visual information such as web screenshots, software interfaces, and charts, and then execute tasks by calling tools. In other words, this “eye” is not for human use but for the agent.The open-source content is quite complete, including model weights, Tokenizer, Prompt Encoding reference implementation, and a minimal PyTorch inference implementation, covering core modules such as the visual encoder, Aligner, DFlash Attention, MoE, and Hyper-Connections. The community responded quickly, with seven quantized versions already available on the Hugging Face page for llama.cpp, LM Studio, and Ollama.There are also notable details in the timeline: On August 21, the model was first launched on the DeepSeek API, at that time only accessible via API without providing weights. Developers could input text and images simultaneously through the API, with images charged by token. Ten days later, the weights were officially opened, making local deployment and secondary development possible. This “sell API first, then open source” approach also reveals DeepSeek’s positioning as primarily an API service.Performance Approaching Claude Opus 4.8, Official Language Remains ModerateFrom the benchmark results, the pure text agent capabilities after adding visual abilities remain largely unaffected: Terminal Bench 2.1 increased from 82.7 to 83.9, and DeepSWE increased from 54.4 to 59.3, even surpassing Opus 4.8’s 58.0. The improvement of the multimodal agent is more significant: ApexBench Pass@1 reached 36.5, Agents’ Last Exam achieved 27.3, exceeding Opus 4.8’s 25.7, and ZeroBench Pass@5 reached 35.0, surpassing the opponent’s 34.0.However, DeepSeek did not claim “comprehensive superiority”: on the NL2Repo project, its score was 57.7, far behind Opus 4.8’s 69.7. The official statement remains moderate, stating only that “multimodal agent capabilities are close to Claude Opus 4.8.” Looking at recent updates, DeepSeek is assembling a complete agent technology stack: the model handles reasoning and tool calls, Harness manages continuous task execution, and Vision allows agents to directly read visual information from the computer. This open-source release is a key move in completing this puzzle.

AI NEWSLatest AI NewsArticleOpenClaw 2.0 Released: Reconstructed the Infrastructure, Supports Multi-User Shared Cloud SessionsPublished in Latest AI NewsTime :Aug 31, 2026Read :3minuteOn August 30, the open source AI project OpenClaw released its largest update to date, OpenClaw 2.0. This update was completed by 933 contributors, with over 16,000 pull requests merged, covering about half of the code submissions in the project’s history. Due to the development workload exceeding the capacity of the original infrastructure, the team spent nearly seven weeks rebuilding the product foundation and delivery process.OpenClaw 2.0 focuses on simplifying installation and reworking the browser experience, and then carries out a deep overhaul of the entire process including messages, memory, skills, models, automation, and security mechanisms. The new version greatly reduces configuration steps, prioritizing the use of existing local API keys, models, or ChatGPT/Claude subscriptions, achieving an extremely simple out-of-the-box experience, with more complex configurations completed through interaction later on.For daily scenarios, the system supports cross-platform connectivity and automated task processing (such as automatically parsing emails and pushing notifications, retrieving information across applications and messages, etc.). In addition, the new version introduces the “Shared Cloud Session” feature, breaking the previous context isolation of individual intelligent agents, allowing team members to collaborate in real-time or fully hand over the AI processing flow.As an open-source project, OpenClaw 2.0 not only improves the deployment efficiency of intelligent agents, but also marks the evolution of AI intelligent agents from single-machine auxiliary tools to collaborative and autonomous productivity infrastructure.

AI NEWSLatest AI NewsArticleGoogle DeepMind releases WeatherNext 3, a meteorological large model that provides hourly updates and 5-kilometer high-precision global forecastsPublished in Latest AI NewsTime :Sep 4, 2026Read :6minuteRecently, Google DeepMind and Google Research jointly released the new global weather AI model WeatherNext3. This model breaks through the inherent bottlenecks of traditional meteorological AI in local terrain resolution and initialization delay by directly accessing real-time geostationary satellite mosaics for hourly initialization and using raw meteorological station observation data for training, achieving a comprehensive upgrade from data architecture to forecast accuracy.In terms of core architecture and input mechanism, WeatherNext3 adopts the same probabilistic functional generation network (FGN) grid Transformer architecture as WeatherNext2 and further expands to multi-resolution output. Its input directly connects to global real-time geostationary satellite mosaics and the HRES analysis data from the European Centre for Medium-Range Weather Forecasts (ECMWF). During the training phase, it extensively absorbs multi-source information including ERA5/HRES-fc0, NASA’s IMERG, meteorological station measured data, and satellite mosaics. Traditional meteorological AI models usually rely on reanalysis grids from numerical weather prediction (NWP), which smooths out local climate changes caused by coastlines, valleys, and mountains. WeatherNext3, however, trains a dedicated output head using raw meteorological station observation data, making its 0.05-degree temperature and dew point forecasts more consistent with actual instrument records rather than relying solely on the model’s own physical atmospheric representation.In terms of resolution and forecast frequency (Cadence), WeatherNext3 can output three levels of forecast results with a single forward pass. The first level is 2-meter air temperature and dew point at a resolution of 0.05 degrees (approximately 5 kilometers) based on meteorological station training. The second level is grid surface data at a resolution of 0.1 degrees (approximately 10 kilometers), including 10-meter and 100-meter wind speed, pressure, sea surface temperature, cloud distribution, solar radiation, and one-hour precipitation. The third level is an atmospheric field covering 13 pressure levels at a resolution of 0.25 degrees (approximately 25 kilometers). Compared to the previous model’s output of 0.25-degree fields every six hours, its accuracy has improved by about five times. In terms of update frequency, the model is initialized 24 times a day. The weather system cycles at UTC 00, 06, 12, and 18 can extend the forecast up to 15 days (360 hours) and include 64 ensemble members, while temporary hourly runs cover a 48-hour window. This hourly rolling refresh based on current satellite observations completely breaks the previous cycle limitations that relied on lagging six-hour traditional numerical forecast analysis, offering significant value for capturing rapidly developing convective weather.In terms of precipitation simulation and support for clean energy variables, WeatherNext3 demonstrates strong business application potential. Precipitation forecasting has always been a historical pain point for global models, often producing unclear fields that miss storm boundaries. This model trains using multi-source data including ECMWF reanalysis, NASA’s IMERG satellite inversion, and Google’s own satellite radar precipitation reanalysis. In the early forecast stage, its Continuous Ranked Probability Score (CRPS) can improve by up to 60% compared to the IMERG benchmark and shows significant improvement compared to meteorological station rain gauges. For the new energy sector, the model can also accurately output 100-meter wind speed at the hub height of wind turbines, complete distributions of low, medium, and high clouds, and all components of solar irradiance. These key data combinations can directly empower power grid operators to accurately predict wind and solar energy generation output.Currently, the forecast data of WeatherNext3 is open for application, and users can access it through BigQuery, Earth Engine, and Cloud Storage. However, the model weights are not open-sourced, and custom inference is still running the previous version of WeatherNext2. As this model is fully implemented, global weather forecasting is accelerating towards a new intelligent era of high precision and high timeliness.

AI NEWSLatest AI NewsArticleAI-Generated Menus Fall into Perfectionist Aesthetics, Food Photos Trigger the Uncanny Valley EffectPublished in Latest AI NewsTime :Sep 4, 2026Read :3minuteGenerative AI is entering the design of restaurant menus, but its highly homogenized “perfect aesthetics” has also raised new issues with visual experience. More and more restaurants are using AI-generated images for food menus, which often exhibit overly symmetrical, smooth, and refined characteristics. Some works even show obvious problems such as abnormal food structures, which consumers describe as having a “uncanny valley effect.“Alex Lisle, Chief Technology Officer of Reality Defender, believes this phenomenon is related to the training mechanism of AI models. LLMs and diffusion models need to learn patterns from massive data, and when models frequently refer to similar-style chain restaurant menus, the generated results tend to become similar; if AI-generated content further enters the training data, it may reinforce this style convergence. Lee Renney, Director of the “Center for Imagining Digital Futures” at Elon University, stated that training data usually tends to be aesthetically pleasing and avoids offensive content, which further pushes images and language towards being “rounded off.“This issue may become more apparent with repeated editing. A user found that after continuously modifying an AI-generated menu 100 times in ChatGPT, the food images gradually deviated from their original form. Researchers from the University of Duisburg-Essen previously found that AI-generated food images may produce a negative perception similar to the “uncanny valley.“Industry insiders point out that AI-generated content is entering the real information environment on a large scale, and the data generated by the model itself may in turn affect subsequent training. How generative AI can avoid aesthetic homogenization and maintain the details and diversity of the real world will become an important issue beyond the model’s capabilities.

AI NEWSLatest AI NewsArticleSmart Ring Manufacturer Oura Has Officially Filed for IPO, 9-Month Revenue Increased to $1.2 BillionPublished in Latest AI NewsTime :Sep 4, 2026Read :4minuteSmart ring manufacturerOura has officially submitted its IPO application. The company filed documents with the U.S. Securities and Exchange Commission (SEC) on Thursday, showing that Oura’s revenue increased from$697 million to $1.2 billionover the past nine months ending June 30. The company previously projected revenue of about $500 million in 2024, $1 billion in 2025, and nearly $2 billion in 2026.As of now, Oura has sold approximately3.6 million smart ringsin the past year, with around5 million paid members, and a 12-month weighted average member retention rate of about 85%. Oura rings are priced between $350 and $400, covering physiological indicators such as heart rate, metabolism, stress, and sleep, and provide continuous health monitoring services through its companion app.Oura was founded in Finland in 2013 and submitted a confidential IPO application in May this year. Earlier market reports indicated that the company planned to raise about3 billion dollarsthrough the IPO, with a potential valuation reaching16 billion dollars; the company’s valuation was approximately 11 billion dollars in October 2025.Oura stated that it will expand from traditional activity and fitness tracking to broader healthcare scenarios, and deepen cooperation with insurance companies, employers, and healthcare providers. The company disclosed that it has accumulated more than50 health indicators and nearly 4.2 billion hours of physiological data, and will use these longitudinal data to train AI and machine learning models to enhance health pattern recognition, personalization, and predictive capabilities.Meanwhile, Oura is facing a proposed class-action lawsuit alleging that it misled consumers regarding the accuracy of its sleep tracking. Oura denies the allegations and states it will defend itself through legal means. This IPO will serve as an important window for observing the evolution of smart wearables from hardware sales to a “hardware + subscription + AI health service” model.