Hugging Face Hub发布的2026年中期生态观察显示，开源模型生态在规模和分布上发生了显著变化。该平台模型库从243万增长至296万，数据集从71.1万增长至100万1。在大规模模型发布上，中国AI实验室表现出明显优势，月度最大模型参数量达2.78万亿，而美国大多数月份的最大模型参数量控制在130B以下1。中国发布的178个20B以上参数的模型中，59%采用Apache 2.0许可，22%采用MIT许可1。
开源模型生态的应用层仍呈现出"小模型为主"的特征。在下载量统计中，小模型占据83%的份额，通过llama.cpp等工具实现本地推理1。100B以上参数的模型仅占全时间下载比例的1%，或在2026年统计中占3%1。Qwen已成为开源生态的基础模型，拥有151,448个衍生模型，日均增速达180至210个新仓库，其总影响力是Meta的2.6倍1。在下载量前25的模型中，13个发布于2022年，而2026年新发布的模型无一进入前251。
在技术生态方面，GGUF库声明增长464%，远超transformers的16%和peft的16%1。代理工具的使用格局也在变化，Claude Code在7月占代理流量的45.4%，但在4月曾达到67.8%1。此外，NVIDIA和AMD各发布了200多个新模型仓库，领先其他机构1。
A mid-year analysis of the open-source model ecosystem reveals a significant shift in global AI development patterns, with Chinese laboratories substantially outpacing their American counterparts in large-scale model releases 1. During 2026, Chinese AI labs released models reaching up to 2.78 trillion parameters monthly, while the majority of American releases remained below 130 billion parameters for five months 1. This disparity underscores evolving competitive dynamics in foundational model development across regions.
The open-source infrastructure continues rapid expansion, with Hugging Face Hub's model library growing from 2.43 million to 2.96 million repositories, while its dataset collection expanded from 711,000 to 1 million 1. Qwen has solidified its position as the dominant baseline model in the open ecosystem, boasting 151,448 derivative repositories with an average of 180 to 210 new repositories added daily 1—representing 2.6 times Meta's cumulative influence 1. Among the 178 models of 20 billion parameters or larger released by Chinese laboratories, 59 percent adopted Apache 2.0 licensing and 22 percent used MIT licensing 1.
Despite the prominence of massive models, practical applications remain concentrated on smaller variants, which account for 83 percent of total downloads 1. Models exceeding 100 billion parameters represent only 1 percent of all-time downloads or 3 percent of 2026 downloads 1, with local inference enabled through tools like llama.cpp 1. Notably, among the top 25 most-downloaded models in 2026, thirteen were released in 2022, while no models debuted that year entered this tier 1. The GGUF library format experienced explosive growth of 464 percent, substantially outpacing transformers and PEFT frameworks, each registering 16 percent growth 1. NVIDIA and AMD each published over 200 new model repositories, leading all other institutions 1.
评论
还没有评论，欢迎留下第一条。