GLM-Z1-Rumination-32B-0414

GLM-Z1-Rumination-32B-0414 is a 32-billion-parameter deep reasoning model released by Zhipu AI on April 14, 2025, as part of the Z1 open-source series. It is built on top of GLM-4-32B-0414 using cold-start techniques and extended reinforcement learning, with targeted training for mathematics, coding, and logic reasoning.

Its distinguishing feature is the rumination mechanism: unlike typical reasoning models that produce a single chain of thought, it engages in longer, reflective deliberation and can integrate search tools mid-reasoning to handle open-ended, complex problems. Zhipu positions it as a deep-research model that autonomously formulates questions, searches for information, constructs analyses, and completes tasks, with benchmark performance reported as comparable to GPT-4o and DeepSeek-V3-0324 on complex tasks.

The model is open-source under the MIT license and shipped alongside GLM-4-32B-0414 and GLM-Z1-9B-0414, a 9B reasoning sibling. The release was accompanied by recommended sampling parameters including a 30,000-token budget for thinking.