最近两三个月 ai大模型动作频出： 国内AI大模型领域呈现出"技术突破与商业落地并驾齐驱"的蓬勃景象。从旗舰模型的性能刷新， 到垂直领域应用的深度拓展，各大厂商在"百模大战"的下半场，正从单纯的参数竞赛转向对实际 价值创造的激烈角逐。 1）通用大模型 2026年1月，阿里巴巴发布了其最新的旗舰推理模型Qwen3-Max-Thinking，创下多项权威评测全 球新纪录。该模型总参数量超过万亿，其核心突破在于采用了一种全新的"测试时扩展"机制，能 够对推理结果进行"经验提取"式的提炼和多轮自我迭代，大幅提升了推理效率与性能。在"人类最 后的测试"（HLE）中，其得分高达58.3，显著超越了GPT-5.2-Thinking和Gemini 3 Pro。更早之前 的2025年12月，百度发布了原生全模态大模型文心大模型5.0，该模型采用超稀疏混合专家架构， 参数规模同样超过2.4万亿，在LMArena文本榜上以1451分的成绩登顶，位居中国第一。 字节跳动旗下的豆包大模型也在2025年12月发布了1.8版本，定向优化了多模态Agent场景，在视 觉推理、通用视觉问答等任务中表现接近全球顶尖水平，其日均tokens调用量 ...

应该喜极而泣？喜极而泣 居然成了避险资产，和以前银行保险一样了 泪流满面 啥时候光模块成了光登 ...

Core Insights - The article discusses the Qwen3-Max-Thinking reasoning model, highlighting its significant performance improvements through technological innovations. Group 1: Core Innovations - The model features an adaptive tool invocation capability, allowing it to autonomously judge and utilize built-in functions like search engines and code interpreters, reducing the likelihood of generating false information [1] - Test-Time Scaling is introduced as an efficient reasoning strategy, enabling the model to extract key insights from past reasoning processes, thus enhancing performance without increasing computational costs [1] Group 2: Performance Metrics - Qwen3-Max-Thinking achieved a score of 58.3 in the "Human Last Exam" (HLE) assessment, surpassing competitors like GPT-5.2-Thinking (45.5) and Gemini 3 Pro (45.8), although it still has room for improvement in specific areas [1] Group 3: Application Benefits - The model has over one trillion parameters and has undergone extensive reinforcement learning, establishing a strong foundation for reasoning and knowledge integration capabilities [2] - It possesses robust agent capabilities, enabling it to autonomously complete complex task processes beyond mere text generation [2] Group 4: Pricing Structure - The pricing for input and output tokens varies based on length, with specific rates outlined for different token ranges, indicating a structured cost model for usage [3] - Comparatively, Qwen's pricing for one million tokens is 3.2 yuan for input and 12.8 yuan for output, positioning it within a competitive pricing landscape [5]

| | | | Others | 2.84T | | --- | --- | | Claude Sonnet 4.5 | 611B | | MiMo-V2-Flash | 462B | | Grok Code Fast 1 | 448B | | Gemini 3 Flash Preview | 437B | | Claude Opus 4.5 | 315B | | DeepSeek V3.2 | 307B | | Gemini 2.5 Flash | 305B | | Gemini 2.5 Flash Lite | 225B | | Grok 4.1 Fast | 224B | | Total | 6.17T | | Weekly Pace | +1.57T | 1）chatgpt 移动端数据环比上行明显。 2）openrouter 数据继续新高！ 3）和上周预测的一样，国产mimo-v2 冲到了第二名！ ...

Group 1 - The article discusses the advancements in space solar power technology, highlighting its potential to surpass terrestrial energy solutions [1] - It mentions the ambitious goal of achieving 30GW of power generation through OpenAI's initiatives over the next five years, indicating a significant shift in energy production paradigms [1] - The tone suggests a sense of overwhelming change in the energy landscape, with space solar power being positioned as a revolutionary force compared to traditional methods [1] Group 2 - There is a commentary on the challenges faced by AI and energy sectors, implying that current energy resources are becoming insufficient [2] - The phrase "电烧干了" suggests a critical view on the sustainability of current energy consumption practices, indicating a potential crisis in energy availability [2]

Core Insights - The article highlights the pervasive inflation logic observed this year, indicating that price increases are widespread across various sectors, particularly in storage, CPU, and cloud leasing services [1] - It mentions that even scarce IDC resources have started to see price hikes, reflecting a broader trend of rising costs in the industry [1] - The phrase "涨价无处不在" (price increases are everywhere) emphasizes the extent of inflationary pressures affecting multiple sectors [1] Industry Summary - Inflation is impacting specific categories such as storage, CPU, and cloud leasing, leading to noticeable price increases [1] - The scarcity of IDC resources is contributing to rising prices, suggesting a tightening supply in the market [1] - The overall sentiment in the industry is that price hikes are becoming a norm, with expectations for continued inflationary trends [1]

数量大，今明两年国内一半需求用自研的芯片，目测大几十万颗，和no1 hw那家比肩了。 互通做得好。磐久，64卡，进发128卡，超结点用UAlink。 用量得以保证，阿里去年10年单日tokens调用超过60万亿！同年11月，字节单日40万亿。昨天发言预计 全球tokens调用还要涨100倍。阿里推理自研用的多。 总结：去年google tpu 外售，阿里就看的眼红，一直紧密跟踪。以为它要外售芯片，结果它直接ipo。 阿里云没有像大家期待的分拆。阿里芯片准备单独分拆ipo。 外号"平头哥"，实际ppu，目前到1.7代，对标h100。 特点： ...

这篇文章很重要‼️。但大家又没看！ 曰 日小分 / // // /// /// 国内的大模型场可以借鉴啦，手机端 os 虚拟机, 云+终端 ai 结合, 那 就 拥有了个性 化, 多好。 质上是 间 非 发 员 AI 编 — 具 Claude Code。Cowork 的大部分功能都由 Claude Code 自行构建, 这使得 AI 既成为 产品本身,也成为开发流程的关键组成部分。 1、它面向普通人 不需要懂编程,编程执行这一步,它已经 内嵌到了工作流本身，也就是Claude code 自从构建。无需编写软件，只需授予 Claude 访问其计算机上某个文件夹的权限,并要求它 执行实际工作,例如整理文件、撰写报告、根 据屏幕截图创建电子表格或清理文档。与普通 聊天工具不同,Cowork 的运行更加自主。您 分配任务后,Claude 会逐步规划并执行,并 随时向您汇报进度。 ai目前把编程变成了英语一类的存在，每 个人都会。或者说ai把编程变成了电,不需 要自己会发电, 你会用就行。 最近我使用 qwen 的内测版做小程序就是这 ------------- 1 . 1 -- 出 a ♡ C 小熊跑的快 赞 今天有 ...

申明：除了阐述事实！什么观点都没有！继前一篇台积电涨价cpu之后的事实补充！没有观点，没意 见，什么观点都没有。几研究gpu的时候，看过amd的，这几年它市场份额扩张的很快。 2025年Q3 AMD x86 CPU总份额达25.6%（环比+1.4pct，同比+1.6pct），首次突破25%；台式机份额 33.6%（环比+1.4pct，同比+5.2pct），首次超1/3；服务器27.8%（环比微增）；移动21.9%（环比 +1.3pct）。 中国区30%，美国30%，剩下中东+欧洲。 去年中国区数据中心业务翻倍了。 份额速览（Mercury Research 2025Q3） 增长趋势 2025Q1-Q3 x86从24%→24.2%→25.6%，台式机从30%→32.2%→33.6%，攀升。 若计入嵌入式/IoT/主机SoC，AMD x86整体份额30.9%（同比+5.9pct）。 • 总x86：AMD 25.6%，Intel 74.4% • 台式机：AMD 33.6%，Intel 66.4% • 服务器：AMD 27.8%，Intel 72.2% • 移动：AMD 21.9%，Intel 78.1% 数据中心， ...

"缝隙越来越大，一片泛着金芒的大海逐渐呈现在了戴里克的眼中，呈现在了科林.伊利亚特的眼中。看 到这样的画面， 浴着温暖的光， 身体一点点"蒸发 "的科林.伊利亚特嘴角微动，露出了一抹不含丝毫阴 霾的笑容和隐隐约约的向往。 光就是一切的意义 白银之城 克莱恩带去了光！ 光是一切的意义。黄昏的巨人，倒在了黎明之前。 赞美愚者！ ...