中国阿里巴巴再度挑战美国在 AI 领域的领先地位

中国科技巨头阿里巴巴发布了其所称“迄今规模最大、能力最强的 AI 模型”，并宣称其性能可媲美美国前沿实验室 Anthropic 和 OpenAI 的顶尖系统，以及 Moonshot AI 的 Kimi K3 等国内竞争对手。阿里巴巴周一在一篇博客文章中表示，正让这款名为 Qwen3.8-Max 的模型面向用户广泛开放。此前外界已预期该模型将发布，因为该公司上个月预览该模型时曾称，它“仅次于 Fable 5”——Anthropic 的旗舰模型。

又一款高能力中国 AI 模型的开放发布，进一步加剧了硅谷与华盛顿在多个议题上的高度紧张局势：如何安全管理 AI 系统，以及如何维持美国相对于中国的技术优势。

阿里巴巴周一公布的内部测试结果，以及其在众包模型对比平台 Arena.AI 上的排名，都表明该模型或许确如其宣称般强大。阿里巴巴的测试显示，Qwen3.8-Max 在基准测试中的表现大体与 Fable 5 持平，且有时超过后者。在 Arena 文本模型排行榜上，Qwen3.8-Max 仅落后于 Fable 5 和 Anthropic Opus 系列的三款模型。在前端编程方面，它仅被两款 Claude Opus 模型和 Kimi K3 超越；而在视觉分析方面，只有 Fable 5 明显领先于它。

阿里巴巴称，Qwen3.8-Max 拥有 2.4 万亿个参数。参数是衡量模型在训练过程中学习到多少配置项的数值指标，模型借此处理数据、识别模式并完成各类任务。尽管参数数量被广泛用作模型性能的简写指标，但数量更大并不总意味着效果更好。Moonshot 的 Kimi K3 模型拥有 2.8 万亿个参数，但大多数美国顶级实验室并不公开这一数字，OpenAI 和 Anthropic 也都未披露其顶级系统的确切参数量。

这家中国公司表示，将在下周发布 Qwen3.8-Max 的权重。权重是决定 AI 如何处理信息的可调整数值。开放权重系统虽然比传统开源软件限制更多，但仍能让开发者获得远多于使用 OpenAI 和 Anthropic 等公司专有产品时的控制权。今年早些时候，阿里巴巴针对更先进模型曾短暂转向专有发布，而 Qwen3.8-Max 标志着该公司重返开放权重发布策略。

开放权重发布已成为中国 AI 行业日益显著的差异化特征，并且在中国逐渐成为常态。Moonshot 上周发布了 Kimi K3 的权重，许多其他顶级 AI 模型也同样采用开放权重形式。北京方面一直将这一战略视为扩大中国在全球 AI 治理中影响力、并推动本土科技龙头企业获得广泛采用的手段。

阿里巴巴此次发布加剧了与中国相关的竞争态势。中国企业似乎正在迅速缩小与美国公司的差距，并在最近几周加快了发布节奏。Qwen3.8-Max 紧随 Kimi K3 之后推出，后者也被视为对美国 AI 主导地位的又一次挑战；而字节跳动和MiniMax也在周五发布了能力强劲的新视频生成模型。

开放性也已被证明是一个充满分歧的问题。在中国模型发布后、美国政府可能收紧对开放工具监管的报道出现之际，美国行业总体上团结起来，支持保留对开放权重模型的访问权限，既将其视为安全上的必要条件，也将其视为维持竞争的手段。这场争论发生之时，闭源模型提供商——尤其是OpenAI和Anthropic——在披露一连串由其自身失控 AI Agent 在无意中实施的网络攻击后，正面临日益严格的审视。来自一名受害者的事件报告表明，原本旨在阻止 AI 模型被恶意使用的严格安全护栏，也会限制这些模型作为防御工具的用途。

Chinese tech giant Alibaba released what it says is its largest and “most capable AI model to date,” claiming performance rivaling the best systems from US frontier labs Anthropic and OpenAI, as well as domestic rivals like Moonshot AI’s Kimi K3. Alibaba said it was making the model, Qwen3.8-Max, widely available to users in a blog post published on Monday. The release had been expected after the company previewed the model last month, when it claimed it was “second only to Fable 5,” Anthropic’s flagship.

The open release of another highly capable Chinese AI model adds to sky-high tensions along multiple fronts in Silicon Valley and Washington over how to safely manage AI systems and retain the US’ technological edge over China.

Results from Alibaba’s own testing shared on Monday suggests it’s as powerful as claimed, as does its ranking on crowdsourced model-comparison platform Arena.AI. Alibaba’s own testing shows the model’s performance to broadly match — and sometimes exceed — that of Fable 5 on benchmark tests. On the Arena text model leaderboard, Qwen3.8-Max trails only Fable 5 and three models in Anthropic’s Opus family. For frontend coding, it is beaten only by two Claude Opus models and Kimi K3, and for visual analysis, only Fable 5 trounces it.

Alibaba says Qwen3.8-Max has 2.4 trillion parameters, a numerical measure of the settings a model learns during training that it uses to process data, recognize patterns, and undertake various tasks. While parameter counts are widely used as a shorthand for model performance, a bigger number does not always mean better. Moonshot’s Kimi K3 model has 2.8 trillion parameters, but most top American labs keep figures private, and neither OpenAI nor Anthropic disclose exact counts for their top systems.

The Chinese company said it will release the weights for Qwen3.8-Max next week. Weights are the adjustable numerical values of an AI that determine how it processes information. Open-weight systems, while more restrictive than traditional open-source software, give developers far more control than they have over proprietary products from companies like OpenAI and Anthropic. Qwen3.8-Max marks a return to open-weight releases for Alibaba after the company briefly pivoted towards proprietary releases for its more advanced models earlier this year.

Open-weight releases have become a growing point of differentiation for China’s AI industry, where they have become the norm. Moonshot released Kimi K3’s last week and many other top AI models are also open-weight. Beijing has championed the strategy as a means of growing China’s influence in global AI governance and encouraging widespread adoption of domestic tech champions.

Alibaba’s release intensifies competition with China, whose firms appear to be rapidly narrowing the gap with US companies and have increased the tempo of releases in recent weeks. Qwen3.8-Max closely follows the release of Kimi K3, viewed as another challenge to American AI dominance, and both ByteDance and MiniMax released capable new video generation models on Friday.

Openness has also proved to be a divisive issue. Amid reports of potential crackdown on open tools in the wake of the Chinese releases, the US industry has largely rallied around preserving access to open-weight models, both as a safety necessity and as means of preserving competition. That debate comes as closed-model providers, notably OpenAI and Anthropic, face increasing scrutiny after revealing a slew of cyberattacks unknowingly perpetrated by their own escaped AI agents. Incident reports from one victim suggest the restrictive safety rails intended to stop nefarious use of AI models also limits their use as defensive tools.