As DeepSeek becomes a focal point on the global stage, its success highlights the growing international presence of Chinese AI companies. Recently, MIT Technology Review published an article titled "Four Chinese AI Startups to Watch Beyond DeepSeek," which identifies and commends four companies, Stepfun, ModelBest, Zhipu, and Infinigence AI, asserting that they have comparable technical capabilities and global competitiveness to DeepSeek.

Founded in 2023 by former Microsoft Senior Vice President Daxin Jiang, Stepfun may be late to the game, but it has quickly gained attention with its eleven foundational AI models. Notably, its large language model, Step-2, has over one trillion parameters and ranks just behind ChatGPT, DeepSeek, Claude and Gemini in the LiveBench benchmark tests.

ModelBest, established by a team from Tsinghua University, focuses on on-device models. Its MiniCPM series, known for its efficiency, has been dubbed the "little cannon." MiniCPM 3.0, with only 4 billion parameters, has achieved performance levels comparable to GPT-3.5 in several benchmark tests, while MiniCPM-o 2.6 has even demonstrated GPT-4o level performance on tablets.

Zhipu, also originating from Tsinghua University, has made significant strides in the development and application of foundational models. Its GLM-4-Plus large language model, trained with high-quality synthetic data, achieves GPT-4-like performance while reducing costs. Meanwhile, its visual model GLM-4V-Plus has made considerable advancements in the capabilities of AI "agents."

Infinigence AI, focused on AI infrastructure, has secured nearly 1 billion yuan in funding within two years. Its "MxN" AI architecture integrates chips from various brands, enabling the efficient collaborative deployment of large model algorithms across diverse chip environments. Its Infini-AI heterogeneous cloud platform and HETHUB mixed training system can shorten model training time by 30%, significantly enhancing computing resource utilization.

Whether it's DeepSeek or these four emerging players, Chinese AI companies are reshaping the global competitive landscape with original technology and engineering breakthroughs.

Source: Lingnan on the Cloud

中国AI走向世界舞台：DeepSeek之外，麻省理工科技点赞“四剑客”

DeepSeek成为全球关注的焦点，其成功也让更多中国AI企业进入国际视野。近日，《麻省理工科技评论》（MIT Technology Review）发布文章《关注DeepSeek之外的四家中国人工智能初创公司》，点名阶跃星辰、面壁智能、智谱和无问芯穹四家企业，认为它们在技术实力和全球竞争力上毫不逊色于DeepSeek。

成立于2023年的阶跃星辰，由微软前高级副总裁姜大昕创立。虽然起步较晚，但已凭借11款基础AI模型崭露头角。其中，大型语言模型Step-2参数超1万亿，在LiveBench基准测试中，仅次于ChatGPT、DeepSeek、Claude和Gemini。  

面壁智能由清华大学团队创立，专注于端侧模型。其MiniCPM系列模型因高效能被称为“小钢炮”。MiniCPM 3.0仅40亿参数，却在多项基准测试中达到GPT-3.5水平；MiniCPM-o 2.6甚至在平板电脑上实现了GPT-4o级别的效果。

同样源自清华大学的智谱，在基础模型研发和应用落地方面取得突破。其GLM-4-Plus大语言模型通过高质量合成数据训练，降低成本的同时，实现与GPT-4媲美的性能；视觉模型GLM-4V-Plus在AI“智能体”能力上也取得重要进展。  

无问芯穹专注于AI基础设施，成立不足两年已获近10亿元融资。其“MxN”AI架构整合不同品牌芯片，实现大模型算法在多元芯片上的高效协同部署。其Infini-AI异构云平台和HETHUB混合训练系统，可将模型训练时长压缩30%，大幅提升算力利用率。  

从DeepSeek到这四家新锐，中国AI企业正以原创技术和工程突破重塑全球竞争格局。

文｜来源于北京日报
译｜洪婷
英文审校｜林佳岱