Tencent creates 16700 working environments for agents: let GLM demonstrate, Qwen follow suit Dynamic Beating AI News: Tencent Hybrid and Fudan University jointly propose WorkForge, which specializes in building complex and verifiable work training environments for AI agents in bulk. The team collected forms, reports, and other materials from real public resources, covering 40 professional fields and 60 file types, and built a total of 16665 training environments. Each workspace contains an average of 42.3 files, which is much more complex than a simulation environment that relies solely on AI to generate materials out of thin air. For example, in financial tasks, the agent needs to analyze stock factor returns in different regions, organize data, compare performance, and finally submit PPT, Excel, and CSV. WorkForge will first extract verifiable facts from the original file, and then automatically generate tasks and scoring criteria. The program is responsible for checking specific numbers and delivery documents, while another set of AI evaluation and analysis content is used to make the training results as reliable as possible. The team uses GLM 5.3, combined with Claude Code and OpenClaw to generate demonstration operation trajectories, and then fine tunes Qwen3.5 with these data. After training, the score of the 35B-A3B basic model in GDPval increased from 45.5 to 73.6, and APEX increased from 5.0 to 21.3. The other 27B model achieved a GDPval of 82.4, slightly higher than Qwen3.5-397B (79.7) with nearly 15 times the number of parameters, but still lower than GPT-5.6 Sol's 90.1. [Original link]