A concept picture of AI city File photo: VCG

Chinese large-language model (LLM) start-ups including DeepSeek and Moonshot AI have rapidly open-sourced their latest models focused on visual understanding and multimodal capabilities, marking fresh momentum in the industry's push to expand the open-source ecosystem and accelerate practical adoption.

The DeepSeek team on Tuesday released a paper titled DeepSeek-OCR 2: Visual Causal Flow on GitHub, an open-source code hosting and collaboration platform, and open-sourced the DeepSeek-OCR 2 model. The model adopts an innovative DeepEncoder V2 approach, enabling artificial intelligence (AI) to dynamically reorder different parts of an image based on its semantic meaning, bringing visual encoding closer to human perception.

Also on Tuesday, Moonshot AI, the Chinese AI start-up behind Kimi, officially released its next-generation open-source model, Kimi K2.5, according to an announcement on the company's official WeChat account.

By integrating visual understanding with reasoning, coding and agent capabilities, Kimi K2.5 lowers barriers to human-AI interaction, allowing users to share images or recordings when text falls short. The model also extends agent functions into everyday office tasks, enabling near-professional use of Word, Excel, PowerPoint and PDF files, the announcement said.

Notably, the model delivered the best overall performance among global open-source models across multiple agent benchmarks, including HLE (Humanity's Last Exam), BrowseComp, and DeepSearchQA, the Star Market Daily reported on Tuesday.

Chinese experts noted that the nation's LLM industry has built a strong research and development (R&D) foundation, with multiple firms capable of developing advanced foundation models. Meanwhile, by leading open-source efforts across data, training and system optimization, the sector has accelerated innovation, adoption and ecosystem-wide progress.

Moreover, market demand and enterprise-level innovation are reinforcing each other in the rapid expansion of open-source multimodal models. 

As more companies push advanced capabilities into open ecosystems, proven experience, code and architectures are being quickly adopted across the industry, accelerating overall development while lowering entry barriers and enabling a broader range of firms to participate in and contribute to the ecosystem, Chen Jing, vice-president of the Technology and Strategy Research Institute, told the Global Times on Tuesday.

The Financial Times on November 26, 2025 reported that China had overtaken the US in the global market for "open" AI models, gaining a crucial edge over how the powerful technology is used around the world. 

Notably, a study by the Massachusetts Institute of Technology and open-source AI start-up Hugging Face found that the total share of downloads of new Chinese-made open models rose to 17 percent in the 2024, according to the Financial Times. The report added that the figure surpassed the 15.8 percent share of downloads from American developers such as Google, Meta and OpenAI — the first time that Chinese groups beat their American counterparts.

"The open-source approach enables rapid adoption of proven code and architectures, accelerating industry-wide development and allowing more firms to join and contribute. Practice shows this path is viable, with high-quality domestic models emerging steadily. While computing power matters, deeper understanding of core principles and sustained innovation are even more critical," according to Chen.