On April 2, Hugging Face, the world’s largest AI open-source community, updated its leaderboard for large AI models. Alibaba’s Qwen2.5-Omni, an end-to-end full-modal model, claimed the top spot, followed by DeepSeek-V3-0324 and SpatialLM-Llama-1B from Manycore Tech. Notably, all three leading models came from Hangzhou-based companies.
Hugging Face’s Open-Source Model Leaderboard
Qwen2.5-Omni: A Breakthrough in Full-Modal AI
Developed by Alibaba’s Tongyi Qianwen, Qwen2.5-Omni can process multiple input types, including text, images, audio, and video, while generating real-time text and performing natural speech synthesis. It set new industry benchmarks in multimodal fusion tasks such as OmniBench.
Unlike closed-source models with massive parameter sizes in the trillions, Qwen2.5-Omni is a compact 7B-parameter model, making it more feasible for deployment on edge devices and various industry applications. Since its release, the model has seen widespread adoption among developers and enterprises, accelerating AI integration into real-world scenarios. To date, Alibaba has open-sourced over 200 models, and the number of Qwen-derived models has surpassed 100,000, surpassing Meta’s Llama series as the world’s largest open-source model family.
DeepSeek-V3-0324: Strong Performance in Complex Reasoning
Ranked second, DeepSeek-V3-0324 has demonstrated exceptional performance in understanding complex instructions, logical reasoning, and knowledge-based tasks. It has also been optimized for Chinese search, writing, and front-end code generation. Prior to this achievement, DeepSeek-V3 had already gained significant traction within the AI community, drawing attention from developers and researchers worldwide.
SpatialLM-Llama-1B: Advancing Spatial Intelligence
In third place, SpatialLM-Llama-1B, developed by Manycore Tech, focuses on spatial understanding. Founded in Hangzhou in 2011 by Huang Xiaohuang, Chen Hang, and Zhu Hao, Manycore Tech specializes in AI-driven spatial intelligence solutions.
SpatialLM-Llama-1B overcomes traditional limitations in geometric and spatial reasoning. For example, given a video, it can generate 3D scene layouts that adhere to real-world physics. This capability has significant potential in architecture, interior design, and virtual scene construction.
Hangzhou’s AI Dominance in Open-Source Models
Back in February, all of the top 10 models on Hugging Face’s leaderboard were fine-tuned derivatives of Alibaba’s Qwen. In Hugging Face’s 2024 open-source model download rankings, Qwen2.5-1.5B-Instruct accounted for 26.6% of total downloads, making it one of the most downloaded open-source models globally.
Hangzhou’s clean sweep of the top three spots in the latest Hugging Face leaderboard is a testament to the city’s growing AI strength. This milestone not only highlights China’s advancements in AI but also helps attract top talent and investment, further diversifying and accelerating AI innovation in the region.
