Huan Qiu Wang·2025-07-30 07:17Core Viewpoint - The "Huanxin Community," a national-level AI open-source platform, was officially launched during the 2025 World Artificial Intelligence Conference, aiming to foster a competitive domestic AI ecosystem through open-source collaboration and innovation [1][2]. Group 1: Platform and Objectives - The "Huanxin Community" focuses on "open-source, collaborative innovation" by providing shared computing resources, model resources, and a development community to lower the barriers to AI technology application [2]. - ZTE Corporation, leveraging its 40 years of communication technology and AI research capabilities, plays a crucial role in the ecosystem's development, committing to deep involvement in model innovation, computing optimization, and practical applications [2][7]. Group 2: Open-source Achievements - ZTE has open-sourced 11 core technological achievements, including 6 self-developed large models and 5 industry-specific datasets, creating a comprehensive technology matrix covering "models, data, and tools" [2][3]. - The NTele-R1-32B-V1 telecom model, trained with only 800 carefully selected samples, achieved a score of 82.5 in the AIME2024 evaluation, surpassing the Qwen3-32B model [3][4]. Group 3: Model Performance - The 3B-Curr-ReFT and 7B-Curr-ReFT models, based on Qwen2.5-VL-Instruct, demonstrated significant performance improvements, with the 3B model achieving an accuracy of 83% in AI2D mathematical reasoning tests, outperforming larger models [4][6]. - The 7B version scored 92.2 in the MathVista evaluation, showing a 33.6 percentage point improvement over the baseline model [4]. Group 4: Industry Impact - The open-sourced datasets cover critical areas such as telecommunications, mathematics, coding, and visual recognition, with the TFCE dataset integrating over 40 years of ZTE's technological expertise [6]. - The collaboration with domestic GPU manufacturers aims to enhance the compatibility of open-source models with domestic chips, achieving a 40% improvement in computing efficiency compared to general solutions [7].