11-19 更新 实时
首页/ AI 大模型/ 正文

阿里巴巴发布 Qwen4-Max 模型 聚焦安全

AI 大模型 来源:Analytics India Magazine 2026-11-16 20:19 0 阅读 机器翻译
阿里巴巴发布 Qwen4-Max 模型 聚焦安全

阿里巴巴推出了 Qwen4-Max,这是其最新的旗舰人工智能(AI)模型。该模型基于全新架构,拥有 2.4 万亿参数和 100 万 Token 的上下文窗口,专为复杂的专业与科学任务而设计。

阿里巴巴表示:“它能够承担广泛的创意与专业工作。”并列举了其在 3D 建模、游戏开发、精密制造、金融、法律、医疗健康以及科学发现等领域的应用。

Qwen4-Max 接替了阿里巴巴于 8 月推出的 Qwen3.8-Max。该公司称,新模型在几乎所有被评估的能力上都优于前代。公司声称,在其 Work360 评估中,Xhigh 和 High 投入等级下的 Token 用量分别下降了约 24% 和 35%。

阿里巴巴还下调了 API 价格,尤其是缓存输入部分。其 Work360 评估中每个任务的平均成本从 4.43 美元降至 1.67 美元,降幅达 62%。此前 Qwen3.8-Max 的发布,是阿里巴巴扩展其 AI 模型组合的又一步。

该公司还强调了安全测试方面的改进。在 Last Secure Code Bench 上,Qwen4-Max 在功能性与安全性两方面的函数级通过率从 Qwen3.8-Max 的 22% 提升至 31.8%。在函数、文件和代码仓库层面满足安全要求的代码占比从 8% 升至 14.7%。

阿里巴巴表示,它加强了对网络安全滥用的防护:保留用于防御性任务(如漏洞检测与修复)的训练数据,同时移除旨在提升攻击性攻击能力的数据。内部测试未发现存在可造成大规模网络安全风险的能力。

针对化学、生物、放射性及核风险,阿里巴巴在训练后阶段引入了专门的安全数据。该公司表示,内部评估发现,该模型识别出了更多潜在有害请求并给出了安全回复,同时并未显著提升用户执行有害任务的能力。

这一安全层面的关注,源于此前 Qwen 模型受到的审视。Cohere 对阿里巴巴一款 Qwen 模型的安全评估,引发了对该模型在测试期间行为的担忧。

阿里巴巴此次发布之际,中国科技公司之间围绕提升 AI 模型性能、同时降低运营成本的竞争正日益激烈。多家中国开发者也已在全球开放权重 AI 市场中崭露头角。

Qwen4-Max 可通过阿里巴巴的 QwenCloud 平台、QwenWork、Qoder 和 QwenStudio 使用。开发者可通过 Token 套餐或 API 访问该模型。

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Alibaba Launches Qwen4-Max AI Model With Focus on Security)

Alibaba has introduced Qwen4-Max, its latest flagship artificial intelligence (AI) model. Built on a new architecture with 2.4 trillion parameters and a 1-million-token context window, the model is designed for complex professional and scientific tasks.

“It can take on a wide range of creative and professional work,” Alibaba said, citing applications in 3D modelling, game development, precision manufacturing, finance, law, healthcare, and scientific discovery.

Qwen4-Max succeeds Qwen3.8-Max, which Alibaba launched in August. The company said the new model outperforms its predecessor across nearly every evaluated capability. In its Work360 evaluation, token usage fell by approximately 24% and 35% at Xhigh and High effort levels, respectively, the company claimed.

Alibaba also reduced API prices, particularly for cached inputs. The average cost per task in its Work360 evaluation fell from $4.43 to $1.67—a 62% reduction. The earlier Qwen3.8-Max launch marked another step in Alibaba's expansion of its AI model portfolio.

The company also highlighted improvements in security testing. On Last Secure Code Bench, Qwen4-Max's function-level pass rate for both functionality and security increased from 22% for Qwen3.8-Max to 31.8%. The share of code meeting security requirements at the function, file, and repository levels rose from 8% to 14.7%.

Alibaba said it had strengthened safeguards against cybersecurity misuse by retaining training data for defensive tasks, such as vulnerability detection and remediation, while removing data intended to improve offensive attack capabilities. Internal testing found no capabilities posing large-scale cybersecurity risks.

For chemical, biological, radiological, and nuclear risks, Alibaba introduced dedicated safety data during post-training. The company said internal evaluations found that the model identified more potentially harmful requests and provided safe responses, without significantly increasing users' ability to carry out harmful tasks.

The security focus comes after scrutiny of earlier Qwen models. A Cohere security evaluation of an Alibaba Qwen model raised concerns about the model's behaviour during testing.

Alibaba's latest release comes amid growing competition among Chinese technology companies to improve AI model performance while reducing operating costs. Several Chinese developers have also gained prominence in the global open-weight AI market.

Qwen4-Max is available through Alibaba's QwenCloud platform, QwenWork, Qoder, and QwenStudio. Developers can access it through a token plan or API.

正文保留英文原文,标题与摘要由机器翻译。

本文由采集器自动同步

原文发布于 Analytics India Magazine(Alibaba Launches Qwen4-Max AI Model With Focus on Security),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗