China AI Breakthrough: Three Chinese Tech Giants Dominate HuggingFace Trending List

Three prominent Chinese technology companies have recently seen their large AI models rise to the top three spots on the trending list of HuggingFace, the world's leading open-source AI community.

The groundbreaking models are DeepSeek-v3, Qunhe Technology's SpatialLM, and Alibaba Cloud's Tongyi Qianwen Qwen2.5-Omni.

Alibaba Cloud reports that Qwen2.5-Omni utilizes the Tongyi team's innovative Thinker-Talker dual-core architecture, which integrates audio-visual technology into its Position Embedding and employs the TMRoPE (Time-aligned Multimodal RoPE) position encoding algorithm.

This Thinker-Talker dual-core design endows Qwen2.5-Omni with capabilities akin to a human "brain" and "vocal cords," creating a unified end-to-end model architecture that enables efficient real-time semantic understanding and speech generation.

On March 27th, Alibaba unveiled and open-sourced its pioneering end-to-end multimodal large model, Tongyi Qianwen Qwen2.5-Omni-7B. This model can simultaneously process diverse inputs, including text, images, audio, and video, and generate real-time text and natural-sounding speech output.

In contrast to closed-source large models boasting hundreds of billions of parameters, Qwen2.5-Omni's compact 7B size paves the way for the widespread industrial application of multimodal large models. Its efficient design even allows for easy deployment and use on mobile devices.

Adding to this achievement, Qunhe Technology's recently released open-source spatial understanding model, SpatialLM, has secured the second position on the HuggingFace trending list.

According to publicly available information, SpatialLM is a proprietary spatial understanding model developed by Qunhe Technology. Remarkably, it can generate physically accurate 3D scene layouts from a single video. Unlike traditional large language models, SpatialLM overcomes limitations in understanding the geometry and spatial relationships of the physical world, promising significant advancements in machine-like spatial cognition and analysis.