根据对美国国会成员的简报分析，中国开源权重AI模型企业已在全球范围内取得领先地位1。截至2025年7月，中国模型的总下载量达到32亿次，是美国同类模型的两倍1。在性能基准测试中，GLM-5.3和Kimi K3等中国顶级模型得分达44-45分，远超Inkling和Nemotron等美国顶级模型的23-26分1。
技术能力方面，中国开源模型目前落后美国闭源模型（OpenAI、Anthropic）约2-5个月，而美国开源模型的差距则为6-9个月1。值得注意的是，中国模型的优势并非仅源于蒸馏技术——即使完全阻止蒸馏，美中差距也仅会扩大1-2个月1。业界普遍认为，中国企业的快速发布周期和针对性优化才是核心竞争力所在1。
学术领域同样反映出这一趋势的加速。截至2026年9月，中文模型在顶级机器学习论文中的被提及率达到38%，超过美国模型的28%，且增速更快1。在推理平台OpenRouter上，中文模型的市场份额从2025年9月的70%上升至80%以上1。包括法律代理应用Harvey、编码工具Cursor以及DoorDash、Airbnb等知名企业已采用中国开源模型1。
Chinese artificial intelligence companies have established a commanding lead in the open-source model landscape, according to an analysis presented to U.S. Congressional members.1 As of July 2025, Chinese models accumulated 3.2 billion downloads, double the volume achieved by American counterparts.1 Leading Chinese models including GLM-5.3 and Kimi K3 scored 44-45 points on performance benchmarks, significantly outpacing top U.S. open-source offerings such as Inkling and Nemotron, which achieved 23-26 points.1
The performance advantage reflects both speed and targeted optimization rather than distillation alone.1 Chinese open-source models lag behind closed-source U.S. systems from OpenAI and Anthropic by an estimated 2-5 months, while American open-source models trail by 6-9 months.1 Even if distillation were completely blocked, the gap would widen by only 1-2 months.1 Academic adoption metrics further demonstrate Chinese dominance: as of September 2026, Chinese-language models were cited in 38 percent of top machine learning papers, surpassing the 28 percent citation rate for U.S. models, with Chinese models gaining share at a faster pace.1
The shift extends across commercial platforms and enterprise applications.1 On OpenRouter, a major inference platform, Chinese models increased their market share from 70 percent in September 2025 to over 80 percent currently.1 Prominent companies including Harvey, Cursor, DoorDash, and Airbnb have already adopted Chinese open-source models in their operations.1
评论
还没有评论，欢迎留下第一条。