Kimi K3 and Baidu Unlimited OCR Sweep Hugging Face Top Two Spots as Chinese Open-Source Models Surge Globally

Chinese AI open-source models are making a sustained global impact, led by two standout releases from Dark Side of the Moon and Baidu.

On the evening of July 27, Dark Side of the Moon (the company behind Kimi) released the full model weights of Kimi K3 as open source. Within tens of minutes, the model shot to the top of Hugging Face's global trending chart, setting a new growth record for the platform. Right behind it was Baidu's Unlimited OCR model, giving Chinese open-source models the top two spots on Hugging Face.

Unlimited OCR is no flash in the pan. It originally launched by topping GitHub Daily Trending (overall and Python), plus Hugging Face's global model trend chart and multimodal trend chart — four number-one positions simultaneously. A month later, the model climbed back to the top of Hugging Face's global chart after being reshared by Turing Award winner Yann LeCun. It now has 19,700+ GitHub stars and 2.65 million Hugging Face downloads.

The model's staying power comes from solving a real pain point: long-document OCR parsing. Traditional OCR models handle books, papers, and reports by processing page by page and stitching results together, causing KV Cache to balloon and inference costs to rise with output length.

Baidu introduced Reference Sliding Window Attention (R-SWA), a mechanism inspired by how humans read and transcribe long documents — staying focused on the source content while keeping only recent generated text as working memory. This lets the model process dozens of pages in a single forward pass with constant KV Cache size, so compute and memory costs stop growing with output length.

Industry observers note that R-SWA's significance extends beyond OCR speed and accuracy. It offers a new paradigm for long-range inference and memory management in large models, pursuing efficient attention mechanisms rather than ever-expanding context windows.

From DeepSeek to Kimi to Unlimited OCR, Chinese open-source AI is transitioning from instant attention upon release to sustained adoption, building genuine long-term value in the global developer ecosystem.