阿里发布新AI模型：更小尺寸，仅1/20参数性能比肩满血DeepSeek-R1

A Chinese open-source AI model is shown to rival top-tier global competitors such as DeepSeek R1, despite its smaller size, representing another step forward in balancing performance and efficiency in AI application.

近日，中国一款新型开源人工智能模型发布，尽管规模较小，实力却能比肩DeepSeek-R1等全球顶级竞品，标志着人工智能应用在平衡性能与效率方面更进一步。

The QwQ-32B, unveiled last Thursday by Alibaba's Qwen team, operates on just 24 GB of video memory with only 32 billion parameters, while DeepSeek's R1 demands 1,600 GB to run its 671 billion parameters, thus realizing a 98-percent reduction.

阿里巴巴通义千问团队上周四推出的QwQ-32B模型，仅需24GB显存且只要320亿参数即可运行，相比DeepSeek-R1模型运行其6710亿参数需要1600GB显存，显存需求减少了98%。

Also, compared to OpenAI's o1-mini and Anthropic's Sonnet 3.7, Qwen's AI model has substantially lower computational requirements.

不仅如此，相较于OpenAI的o1-mini模型和Anthropic的Sonnet3.7模型，QwQ-32B的计算需求也大幅下降。

Kyle Corbitt, a former Google engineer, published his testing results on social media platform X, showing that "the smaller, open-weight model can match state-of-the-art reasoning performance."

前谷歌工程师凯尔·科尔比特（Kyle Corbitt）在社交媒体平台X上公布了他的测试结果，结果显示“这款参数更少的开源模型能够达到最先进的推理性能水平”。

According to Corbitt's team, QwQ-32B achieved the second-highest score in a deductive reasoning benchmark via a method called reinforcement learning (RL), outperforming R1, o1 and o3-mini, while nearly matching Sonnet 3.7's performance at an inference cost more than 100-fold lower than that required by Sonnet 3.7.

据科尔比特团队透露，在演绎推理基准测试中，QwQ-32B运用强化学习（RL）方法取得了第二高的分数，超越了R1、o1和o3-mini，且推理成本较Sonnet 3.7降低100多倍，性能却与之相当。

"AI isn't just getting smarter, it's learning how to evolve," commented Shashank Yadav, CEO of Fraction AI. "QwQ-32B proves that reinforcement learning can out-compete brute-force scaling."

Fraction AI首席执行官沙尚克·亚达夫（Shashank Yadav）对此评论道：“人工智能不仅变得更智能了，它还在学习如何进化。QwQ-32B证明了，强化学习可以超越传统单纯依靠扩大规模的方法。”

"We found RL training enhances performance, particularly in math and coding tasks. Its expansion can enable medium-sized models to match large MoE models' performance," read Qwen's blog article on Github.

通义千问团队在GitHub博客中也指出，强化学习训练对模型在数学和编程任务中的表现提升尤为显著，有望推动中型模型达到大型混合专家（MoE）模型的性能层级。

Qwen's new model is expected to enhance the feasibility of local operations for generative AI products on computers and even mobile devices in the future.

未来，该模型有望提高生成式AI产品在电脑乃至移动设备上进行本地运行的可行性。

Awni Hannun, a computer scientist at Apple, has run QwQ-32B on the Apple computer powered by its M4 Max chip, and it appears to be "running nicely."

苹果公司计算机科学家奥尼·汉农（Awni Hannun）在搭载M4 Max芯片的苹果电脑上运行了QwQ-32B，结果发现运行非常流畅。

China's national supercomputing internet platform last Saturday announced the launch of the API interface service for QwQ-32B. In addition, Biren Technology, a Shanghai-based GPU chip designer, announced Sunday that it has launched an all-in-one machine capable of running this model.

中国国家超算互联网平台9日宣布推出QwQ-32B的应用程序编程接口（API）服务。此外，上海的GPU芯片设计公司壁仞科技周日宣布，已推出一款能够运行QwQ-32B模型的一体机。

QwQ-32B is freely accessible as an open-source model that anyone can run, following DeepSeek's path of facilitating wider application of AI technologies worldwide and contributing China's wisdom to the world.

QwQ-32B作为一款开源模型可供任何人自由使用和运行，它也延续了DeepSeek推动AI技术全球普及的路线，为世界贡献了中国智慧。

Alibaba also recently open-sourced its AI video-generating model Wan2.1, which is available for download on Alibaba Cloud's AI model community, Model Scope and the collaborative AI platform Hugging Face.

此外，阿里巴巴最近还开源了其人工智能视频生成模型通义万相2.1（Wan2.1），该模型可在阿里云的人工智能模型社区魔搭（Model Scope）以及人工智能协作平台Hugging Face上下载。

The e-commerce and cloud-computing giant has announced a plan to invest more than 380 billion yuan (about $52.97 billion) in building cloud and AI hardware infrastructure over the next three years.

作为电子商务与云计算领域的领军企业，阿里巴巴还宣布了一项重大投资计划。未来三年，阿里巴巴将投入超过 3800 亿元人民币（约合 529.7 亿美元），用于构建云和人工智能硬件基础设施。