AutoGLM Rumination

AutoGLM Rumination (AutoGLM沉思) is an agent product launched by Zhipu AI at the Zhongguancun Forum in late March 2025, combining deep research with actual task execution — what Qiming Venture Partners described as pushing agents into a "think while doing" stage, billed as the world's first agent to merge both capabilities . Under the hood it runs on Zhipu's "rumination model," trained via reinforcement learning to self-criticize and reflect, stitched together from GLM-4's general capability, GLM-Z1's reflection, GLM-Z1-Rumination's deliberation, and AutoGLM's execution; the core models were open-sourced on April 14, 2025 . In practice, as 十字路口Crossing tested it, the agent autonomously browses the web — pulling from sources like arXiv and Xiaohongshu, consulting 84 references in one demo — and outputs long cited research reports, all free and unlimited on Zhipu Qingyan's web, PC, and mobile apps . Not everyone was impressed: 葬AI dismissed it as "clearly a me-too Agent product," complaining its overlong chains of thought lacked even basic auto-collapse in the UI .