字号·· | 护眼
南华早报

阿里巴巴:新旗舰AI模型将任务成本降低60%,同时收紧网络安全防护Alibaba says new flagship AI model cuts task costs by 60% while tightening cyber guardrails

点「原文对照」整页切到原文,或双击某段只看那段的原文。

阿里巴巴集团推出了迄今为止最先进的人工智能模型 Qwen4-Max,并配备了严格的网络安全护栏,这家科技巨头在 AI 安全与持续坚持开源 AI 的承诺之间寻求平衡,此举正值全球对 AI 赋能的网络攻击担忧日益加剧之际。

Alibaba Group Holding has launched its most advanced artificial intelligence model to date, Qwen4-Max, with strict cybersecurity guardrails, as the tech giant balances AI safety with a continued commitment to open-source AI amid growing global concerns over AI-enabled cyberattacks.

阿里巴巴表示,该模型的发布凸显了中国 AI 行业对安全的日益重视,公司有意过滤了模型的训练数据,以移除攻击性网络攻击能力,同时保留了用于网络防御的能力,如漏洞检测。

The release highlights a growing focus on safety in China’s AI industry as Alibaba said it intentionally filtered the model’s training data to remove offensive cyberattack capabilities while retaining capabilities useful for cyber defence such as vulnerability detection.

这家总部位于杭州的公司周一表示,还引入了专用数据集和训练后增强技术,以减轻化学、生物、放射性和核(CBRN)风险。

The Hangzhou-based company also introduced dedicated data sets and post-training enhancements to mitigate chemical, biological, radiological and nuclear (CBRN) risks, it said on Monday.

公司还强调了该模型在科学研究方面的改进,包括完成了一次持续 48 小时的运行,发现了一种经实验验证的下一代半导体芯片冷却剂分子,以及对阿尔茨海默病的自主研究。

The company also highlighted the model’s improvements in scientific research, including the completion of a continuous 48-hour run to discover an experimentally validated coolant molecule for next-generation semiconductor chips, as well as autonomous research on Alzheimer’s disease.

关于中国公司青睐的开放权重模型,还是美国竞争对手如 OpenAI 和 Anthropic 等主张的闭源模型,哪种更可能增加 CBRN 风险,目前存在激烈争论。

There is heated debate about whether open-weight models, which are favoured by Chinese companies, could heighten CBRN risks more than the closed-source models championed by their US rivals such as OpenAI and Anthropic.

一些专家警告称,AI 系统最终可能变得足够强大,使恶意行为者或国家能够开发 CBRN 武器。

Some experts have warned that AI systems could eventually be powerful enough to enable malicious actors or states to develop CBRN weapons.

虽然 Qwen4-Max 是闭源模型,但阿里巴巴重申了其开源战略,表示将很快向全球开发者发布 Qwen4 系列的小规模开放权重版本。

While Qwen4-Max is a closed model, Alibaba reaffirmed its open-source strategy by noting that smaller-scale, open-weight versions of the Qwen4 series will be released to global developers soon.

阿里巴巴的 Max 系列模型一直是闭源的,直到 Qwen 3.8 于 8 月开源。Max 系列也是该公司参数规模最大的模型。不过,公司强调了其在让最新模型为用户提供服务时更具成本效益方面取得的成果。

Alibaba’s Max series of models had always been closed-source until Qwen 3.8, which was made available as open source in August. The Max series is also the company’s largest model by parameter size. However, the company highlighted its achievements in making its latest model more cost-efficient when served to users.

Qwen4-Max 拥有 2.4 万亿参数,规模与其前代 Qwen3.8-Max 相同,但每个 token 仅激活 710 亿参数 —— 比前代减少 25%。

Qwen4-Max boasts 2.4 trillion parameters, the same size as its previous Qwen3.8-Max, though it activates only 71 billion parameters per token – 25 per cent less than its predecessor.

据阿里巴巴称,引用名为 Work360 的内部基准测试(评估模型在各类工作任务上的表现),这使其能在提供更强推理能力的同时降低推理成本,将每项任务的平均成本降低 60%。

That allows it to deliver superior reasoning capabilities while bringing down inference costs, cutting the average cost per task by 60 per cent, according to Alibaba, citing an internal benchmark called Work360, which evaluates model performance on various work tasks.

该模型目前可通过 QwenCloud 使用,输入定价为每 100 万 token 0.075 美元。

The model is currently available via QwenCloud, with input pricing set at US$0.075 per 1 million tokens.

阿里巴巴的 AI 努力重点在于抢占企业 AI 市场并培育开源社区,以超越 DeepSeek 和 Z.ai 等本土竞争对手,以及美国竞争对手。

Alibaba’s AI efforts have focused heavily on capturing the enterprise AI market and cultivating an open-source community to outpace local rivals like DeepSeek and Z.ai, as well as US competitors.

阿里巴巴 Qwen 大语言模型(LLM)团队负责人刘大奕衡 9 月展示了路线图,概述了 Qwen 4、Qwen 4.5 和 Qwen 5 系列的计划,并于本月初在乌镇世界互联网大会上强调团队对开源社区的承诺。

Liu Dayiheng, head of Alibaba’s Qwen large language model (LLM) team, presented the road map outlining plans for the Qwen 4, Qwen 4.5 and Qwen 5 series in September, while stressing the team’s commitment to the open-source community earlier this month at the World Internet Conference in Wuzhen.

公司表示,其下一代模型 Qwen 5 将扩展至 5 万亿至 10 万亿参数。阿里巴巴拥有《南华早报》。

The company has said that its next generation model, Qwen 5, would be scaled up to between 5 trillion and 10 trillion parameters. Alibaba owns the South China Morning Post.

阅读媒体原文 ↗