阿里巴巴发布迄今“最强大”的AI模型

qimuai 发布于 阅读:16 一手编译

阿里巴巴发布迄今“最强大”的AI模型

内容来源:https://aibusiness.com/generative-ai/alibaba-unveils-most-powerful-ai-model-yet

内容总结:

阿里巴巴于本周一正式发布其迄今规模最大、功能最强的AI大模型——通义千问3.8-Max。据阿里云官网介绍,该模型采用开放权重设计,参数规模高达2.4万亿,支持最高100万token的上下文窗口。其能力覆盖代码编写、科学研究、长周期任务执行及视觉智能等多项领域,阿里方面表示,该模型可在极少人工干预下,连续数周自主完成编程任务。

阿里在声明中指出,通义千问3.8-Max专为处理复杂的现实世界工作负载而设计,可广泛应用于应用设计、法律文件审阅、体育数据分析、金融研究、菜品创意开发、康复进度可视化及建筑3D建模等多元场景。阿里同时公布的模型评测排名显示,通义千问3.8-Max在多项指标上与Anthropic旗下Fable 5模型表现相当,部分领域甚至更优。

该模型即日起已通过阿里云百炼平台向全球开发者开放API接口,模型权重计划于下周正式全面开源。

此次发布正值中国科技企业竞相研发高性能且成本可控AI模型的激烈竞争期。约一个月前,国内同行月之暗面已推出其最新模型Kimi K3,该模型参数规模达2.8万亿,被称为中国迄今最大的AI大模型,同时也是全球首个开放的3T级参数模型,主打长周期编程、知识工作及深度推理等前沿智能任务。更早前,DeepSeek于上周五发布V4-Flash模型,早期基准测试数据显示,该模型有望成为全球运行成本最低的AI模型之一。

与此形成对比的是,美国主流模型开发商如OpenAI、Anthropic和谷歌等,多采用闭源专有模式;而中国厂商则几乎清一色选择开放权重路线,允许公众获取并修改模型。这一系列密集发布,凸显了中国AI企业以高性价比和开源策略抢占全球AI产业高地的战略意图。

中文翻译:

本内容由谷歌云赞助
选择你的首批生成式AI应用场景
要开始使用生成式AI,首先要聚焦那些能够改善人类获取信息体验的领域。
Qwen3.8-Max是中国科技公司竞相开发强大且价格实惠的AI模型的最新成果。
阿里巴巴周一发布了Qwen3.8-Max,称这是其迄今规模最大、功能最强的AI模型。
在阿里云网站上发布的一篇帖子中,这家中国综合企业表示,这款开放权重模型拥有2.4万亿参数,支持高达100万token的上下文窗口。其能力还包括编程、研究、长周期任务和视觉智能,阿里巴巴称该模型可以在几乎没有人工输入的情况下自主编程数周。
“Qwen3.8-Max旨在处理复杂的现实世界工作负载,涵盖应用设计、法律文件审阅、体育分析、金融研究、烹饪概念开发、康复进展可视化和建筑3D建模等领域,”该公司在一份声明中表示。
阿里巴巴分享的AI模型排名显示,Qwen3.8-Max的得分与Anthropic的Fable 5相当,有时甚至更高。
该模型计划于下周大规模发布,目前全球开发者已可通过阿里云百炼平台上的API使用,模型权重也计划于下周发布。
这距离国内竞争对手月之暗面发布其最新模型Kimi K3大约一个月,后者拥有2.8万亿参数,是中国最大的AI模型。
月之暗面还将其描述为全球首个开放的3T级(三万亿参数)模型,专为长周期编程、知识工作和推理等领域的前沿智能而设计。
更近期,DeepSeek于周五发布了其V4-Flash模型,早期基准数据表明它可能是全球运行成本最低的AI模型之一。
这一系列发布的背景是,中国科技公司正竞相开发越来越智能且运行成本不至于高得离谱的AI模型。与包括OpenAI、Anthropic和谷歌在内的美国模型开发商相比,中国厂商几乎清一色开发开放权重模型(而非闭源模型)。开放模型允许公众访问并进行修改,而闭源的专有模型则不允许。

英文来源:

Sponsored by Google Cloud
Choosing Your First Generative AI Use Cases
To get started with generative AI, first focus on areas that can improve human experiences with information.
Qwen3.8-Max is the latest in a series of launches from Chinese tech firms racing to develop powerful, affordable AI models.
Alibaba on Monday introduced Qwen3.8-Max, which it says is its largest and most powerful AI model to date.
In a post on the AlibabaCloud website, the Chinese conglomerate said the open-weight model boasts 2.4 trillion parameters and supports a context window of up to 1 million tokens. Capabilities also include coding, research, long-horizon tasks and visual intelligence, with Alibaba saying the model can code autonomously for weeks with little human input.
“Qwen3.8-Max is engineered to manage intricate, real-world workloads, spanning application design, legal document review, sports analytics, financial research, culinary concept development, rehabilitation progress visualization, and architectural 3D modeling,” the company said in a statement.
Alibaba shared AI model rankings that showed Qwen3.8-Max delivered comparable, and sometimes better, scores than Anthropic’s Fable 5.
The model, which is scheduled for wide-scale release next week, is now accessible using APIs on the Alibaba Cloud Model Studio for global developers, with model weights scheduled for release next week.
It comes about a month after domestic rival Moonshot AI released its latest model, Kimi K3, which boasts 2.8 trillion parameters, making it China’s largest AI model.
Moonshot also described it as the world's first open 3T-class (three trillion-parameter) model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning.
More recently, DeepSeek released its V4-Flash model on Friday, with early benchmark data suggesting it could be one of the cheapest AI models to run globally.
The series of launches comes as Chinese tech companies race to develop increasingly smart AI models that aren’t prohibitively expensive to run. In contrast to U.S. model developers, including OpenAI, Anthropic and Google, Chinese vendors are almost exclusively developing open-weight models (as opposed to closed source). Open models allow public access for modifications, while closed, proprietary models do not.

商业视角看AI

文章目录


    扫描二维码,在手机上阅读