开源AI模型研究资料清单发布

一份涵盖开源AI模型的研究资料汇编已于2026年9月13日发布。1该清单由AI研究者为政策相关写作而编制，汇集了关于开源模型定义、商业战略、美中竞争态势和技术细节的学术文献与政策文献。1

根据该资料汇编的梳理，开源模型与闭源模型之间的性能差距已缩小至4-6个月。1自2024年以来，处于领先地位的开源模型均来自中国实验室，DeepSeek V4 Flash等模型已进入帕累托成本前沿。1这一态势引发了西方政策制定者的关注，DoorDash、Airbnb和Apple等企业因使用中国模型而受到美国议员调查。1

模型蒸馏已成为2026年开源模型领域最重要的技术与政策争议焦点。1中国公司通过蒸馏及其他技术手段大规模使用Anthropic产品，frontier模型API被发现存在系统性提取推理轨迹的漏洞。1商业层面，Perplexity于2025年1月28日迅速采用DeepSeek R1以降低成本，Thomson Reuters也于2026年8月24日采用Qwen模型从而转离Claude。1

A comprehensive reading list examining open-source artificial intelligence models has been compiled by AI researchers for policy-related writing purposes.1 The curated bibliography, last updated on September 13, 2026, addresses critical dimensions of the open-source AI landscape including model definitions, commercial strategies, US-China competitive dynamics, and technical specifications.1

The performance gap between open-source and proprietary AI models has narrowed to 4-6 months, according to the research compilation.1 Since approximately 2024, leading open-source models have originated from Chinese laboratories, with systems like DeepSeek V4 Flash positioned at the Pareto cost frontier.1 This shift has prompted investigation by US lawmakers: DoorDash (July 31, 2026), Airbnb (April 29, 2026), and Apple (May 17, 2025) faced congressional scrutiny over their use of Chinese AI models.1 Perplexity rapidly adopted DeepSeek R1 to reduce operational costs on January 28, 2025.1 Thomson Reuters migrated from Claude to Qwen-based systems on August 24, 2026.1

Model distillation emerged as the most significant technical and policy debate within open-source AI development during 2026.1 Research indicates that Chinese companies have employed distillation techniques and other methods to systematically extract value from Anthropic's frontier model APIs, exploiting documented vulnerabilities that enable the extraction of reasoning traces, as detailed in 2026 academic papers.1