Kimi K3 and Unlimited OCR Top Global Charts, Chinese Open-Source Models Sweep Overseas

On July 27, Moonshot AI open-sourced the complete model weights of Kimi K3, which rapidly climbed to the top of Hugging Face's overall trend list within minutes. Immediately after, Baidu's Unlimited OCR model, which had been leading the list for days, secured the second spot. These two Chinese open-source models dominated the Hugging Face rankings, becoming a phenomenon. Unlimited OCR achieved an extraordinary feat: within five days of its release, it garnered over 10,000 GitHub stars and topped both the GitHub Daily Trending overall and Python lists, as well as Hugging Face's global model and multimodal model trend lists, achieving a quadruple first place. One month later, Unlimited OCR regained the top spot on Hugging Face's global model trend list after being retweeted by Turing Award winner Yann LeCun. Its GitHub stars have exceeded 19,700 and Hugging Face downloads reached 2.65 million. The popularity of Unlimited OCR stems from its breakthrough in long-document parsing. Traditional OCR models handle long documents by parsing page by page and concatenating results, causing KV Cache to grow during decoding, increasing inference speed and memory costs. Baidu proposed the Reference Sliding Window Attention (R-SWA) mechanism, which maintains attention to the original document while only retaining the most recent generated content as working memory, allowing continuous parsing of dozens of pages in a single forward pass. This keeps KV Cache constant, reducing computational cost and memory usage. Industry experts note that this innovation not only speeds up OCR but also offers new approaches for long-context reasoning and memory management. Unlimited OCR's sustained developer interest reflects that Chinese AI open-source models are moving from initial buzz to continuous adoption.