China's first trusted data space pilot, launched in 2025, has reached a milestone covering 13 cities, 22 industries and 28 enterprises. It has generated 457 high-quality datasets and 136 vertical AI models, enabling scenarios such as multimodal logistics, mutual recognition of carbon footprint accounting and new materials R&D.
Vertical Model
垂直领域模型
Large models deeply customized for specific industries or scenarios, emphasizing domain knowledge and task fit.
Curated Intelligence
Developers discovered code references to a dedicated 'Cyber' security model on DeepSeek's website, featuring personal and enterprise versions. Currently in limited beta testing, access requires strict identity verification and qualification reviews for API keys, with no official announcement from DeepSeek yet.
Zhipu AI has released GLM-Legal, a specialized open-source model that ranks first among open-weight models on both Harvey LAB and LegalBench benchmarks. Optimized for Chinese and cross-border legal tasks, it demonstrates strong capabilities in contract review, evidence analysis, and compliance research, signaling improved precision in vertical domain applications.
Preferred Networks (PFN) released the fully self-developed PLaMo 3 Translate 31B model, supporting 53 languages and online meeting translation. The model outperformed OpenAI's GPT-6.1 Sol in four benchmarks and reduced inference costs to just 14 JPY per 100k characters through efficient Japanese processing.
LifeSciBench has been introduced as an expert-authored and reviewed benchmark designed to evaluate how AI systems handle real-world life science research tasks and decisions. This tool fills a gap in vertical domain evaluation by providing standardized metrics for assessing large models' utility in scientific discovery.
OpenAI and Retro Bio utilized the specialized GPT-4b micro model to engineer more effective proteins for stem cell therapy and longevity research. This case highlights the acceleration potential of vertical-domain small models in bio-computation, though specific performance metrics and technical details are lacking.