Co-built by 31 Chinese mainland universities and six universities from Hong Kong and Macao, the Shenzhen Loop Area Institute (SLAI) has yielded impressive results in just one year: completing full-parameter post-training of a 1.6-trillion-parameter large model powered by a domestic Ascend 910C computing cluster; taking first place among more than 1,300 competing teams at the 2026 ITU Telco Troubleshooting Agentic Challenge with a team of doctoral students enrolled for less than a year; and generating over 100 research deliverables. Yet its report card is defined not merely by numbers, but by a pioneering operational paradigm.

How does this new institute recruit students? For Zhi-Quan Luo, Vice President of The Chinese University of Hong Kong, Shenzhen (CUHK-Shenzhen) and Executive Dean of the institute, the answer lies in "selection bootcamps, various AI competitions, and specialized recruitment tracks," where candidates are assessed through real-world tasks and only those with proven potential are admitted. Across two cohorts, a total of 509 Ph.D. students have been drawn from the 37 partner universities, spanning 65 academic tracks, with students from Hong Kong and Macao accounting for 48% of the student body; faculty members are recruited globally.

Evaluation criteria are also shifting: academic papers are no longer the sole yardstick. Assessments from both the academic community and industry are being integrated into the evaluation system, alongside the rollout of a dual-track mentorship model involving industry advisors. Moving the classroom to the industrial frontline is not about simply replicating corporate R&D workflows; rather, it focuses on critical links such as compute adaptation and foundation model R&D, enabling students to tackle high-impact problems through hands-on training.

Junchao Huang, a Ph.D. student jointly trained with CUHK-Shenzhen, is a lead developer of SolarWM, a streaming interactive world model. "From day one, this project was geared toward industrial deployment rather than remaining a mere academic demonstration," he said. His team is dedicated to providing a foundational framework that lowers the barrier to building world models. "Traditional academic research groups rarely have access to large-scale compute and storage resources, whereas the Shenzhen Loop Area Institute can provide them." Over the past year, student teams have taken on pivotal roles across multiple frontier research initiatives.

In 2025, the world's first Ascend 384 SuperNode for the education and research sector was deployed at the institute. Immediately following the deployment, the institute assembled a project team for its AI training platform, partnering with Harbin Institute of Technology, Shenzhen, the Shenzhen Research Institute of Big Data (SRIBD), and relevant teams at Huawei to conduct joint research in synergy with the Shenzhen Smart City AI Computing Platform. In June this year, leveraging the domestic Ascend 910C computing cluster, the joint research team completed full-parameter post-training for the 1.6-trillion-parameter DeepSeek-V4 series large models, substantially boosting training efficiency while reducing training costs.

Jinchao Xu, Director of the Center for AI for Science and Engineering at the Shenzhen Loop Area Institute and a mathematician by training, is leading a team in collaboration with multiple Shenzhen institutions and hospitals to detect cancer cells in blood by integrating artificial intelligence, scientific computing, and microfluidic chip technology, striving to drive down screening costs. The team uses microfluidic chips to enrich large cells in blood samples, followed by microscopic imaging coupled with AI-powered image recognition to identify cancer cells—achieving a technological breakthrough spanning from cell capture to intelligent identification. They have also independently developed BioMod, a single-cell foundation model, and the project has already signed cooperation agreements with several Class-A tertiary hospitals.

37所高校共建一所“新学院”，一年能跑出什么？

深圳河套学院由31所内地高校及6所港澳高校共建，短短一年，成果丰硕：依托昇腾910C国产算力集群完成1.6万亿参数大模型全参数后训练；入学未满一年的博士生团队在2026年ITU全球电信AI Agent大赛上，与1300多支队伍同台竞技并最终夺冠；累计形成科研成果100余项。答卷不只是数字，更是一套新打法。

新学院怎么招学生？香港中文大学（深圳）副校长、执行院长罗智泉的答案是“选拔营、各类AI大赛加上专项计划”，学生需要在真实任务里接受考察，有潜力的才进得来。据介绍，两届共509名博士生来自37所共建高校，覆盖65个专业方向，港澳生占比为48%；师资面向全球招聘。

评价口径也在变：论文不再是唯一的评价标尺，学术共同体的评价和产业界的评价都将纳入考核体系，产业导师参与的双轨制培养也正在铺开。把课堂“搬”进产业一线，并非简单复现企业的科研业务，而是聚焦算力适配、基础模型研发等关键环节，让学生在训练中解决重要问题。

与香港中文大学（深圳）联合培养的博士生黄俊超，是流式交互世界模型SolarWM的主要研发者。“这个项目一开始就是面向产业落地的，而非仅仅停留在学术演示。”他说，团队致力于提供基础框架、降低世界模型的搭建门槛，“传统课题组很难拿到大规模算力、存储资源，但深圳河套学院可以提供”。一年间，多项前沿攻关项目由学生团队承担关键任务。

2025年，全球教育科研领域首套昇腾384超节点落地学院。学院随即组建AI训练平台项目团队，联合哈尔滨工业大学（深圳）、深圳市大数据研究院、华为有关团队，协同深智城AI算力平台开展联合攻关。今年6月，联合研究团队依托昇腾910C国产算力集群，完成1.6万亿参数大模型DeepSeek—V4系列全参数后训练，大幅提升模型训练效率、降低训练成本。

数学背景的科学与工程智能中心主任许进超，带团队联合深圳多家单位和医院，融合人工智能、科学计算与微流控芯片技术检测血液中的癌细胞，努力降低筛查成本。据介绍，团队用芯片富集血液中的大细胞，再以显微成像加AI图像识别辨认癌细胞，实现从细胞捕获到智能识别的技术突破；还自研单细胞大模型BioMod，目前该项目已与多家三甲医院签约。

文｜记者 沈婷婷
图｜学校供
翻译｜陈冠臻
审校｜曾敏