中国科学院深圳先进技术研究院喻学锋研究团队在《自然-机器智能》期刊发布了材料科学领域首个轻量化双模型协作智能体MatBrain1。该智能体由参数量分别为30亿和14亿的Mat-R1和Mat-T1两个轻量化模型组成1，在材料研究中的预测综合准确率相比GPT-5、DeepSeek-R1等大模型提升66%，部署成本降低95%1。
MatBrain已开发成公众可下载的体验版软件1。这一成果已入选2025年度中国科学院网信工作十大进展1。此外，该研究团队获批首批国家级制造业中试平台，全国共21家入选，其中材料领域仅2家1。
A research team led by Yu Xuefeng at the Chinese Academy of Sciences' Shenzhen Institute of Advanced Technology has released MatBrain, described as the first lightweight dual-model collaborative intelligent agent for materials science research. 12 The system, published in Nature Machine Intelligence, comprises two compact models designated Mat-R1 (with 3 billion parameters) and Mat-T1 (14 billion parameters). 1
The MatBrain agent demonstrates a 66 percent improvement in prediction accuracy compared to large models including GPT-5 and DeepSeek-R1, while reducing deployment costs by 95 percent relative to comparable large-scale models. 12 The team has developed a downloadable demonstration version of the software for public access. 1 The core technology behind the system—described as a dual-model collaborative lightweight materials science intelligent research method and system—has been patented. 1 The research achievement has been selected as one of the ten major advances in the Chinese Academy of Sciences' information and communications work for 2025. 1 Additionally, the institute has been granted approval as one of only 21 national-level manufacturing incubation platforms nationwide, and one of merely two in the materials field. 1
评论
还没有评论，欢迎留下第一条。