Z.ai, formerly known as Zhipu, has open-sourced its GLM AI model series, including its base, reasoning, and rumination models, all distributed under the MIT license.
The GLM-4-32B-0414 base model is pre-trained on 15 terabytes of data, including synthetic data focused on reasoning. It is designed to support a range of tasks, including code generation, report writing, and question answering. Additionally, it has been post-trained to integrate human dialogue scenarios, as well as reinforcement learning techniques aimed at improving performance in structured tasks such as function calling and engineering code generation.
The reasoning model, GLM-Z1-32B-0414, builds on the base model with additional reinforcement learning and optimization techniques. It is tailored for tasks requiring advanced problem-solving, particularly in mathematics, coding, and logic. Z.ai reports that it performs competitively with significantly larger models such as DeepSeek-R1 in specific benchmarks.
The GLM-Z1-9B-0414 model offers similar capabilities at a smaller scale. Trained with the same methods as the larger models, it aims to deliver strong performance for users in resource-limited environments.
Z.ai has also released GLM-Z1-Rumination-32B-0414, a model designed for complex, open-ended problem solving. It uses multi-step reasoning and can incorporate external search tools to complete tasks such as research writing and in-depth analysis. It introduces mechanisms for autonomous question formulation, information gathering, and result synthesis.
These models are now accessible for free through the newly launched Z.ai website, which sports a ChatGPT-like conversational interface. It currently allows users to select between GLM-4-32B, Z1-32B, and Z1-Rumination models. According to the press release, all models are available for free commercial and research use under the MIT license.
Z.ai is Zhipu’s new name for the international market. This is a part of the company’s efforts to ramp up its global presence ahead of its initial public offering, the first Chinese artificial intelligence startup to do so. Currently, the company is preparing for an IPO as early as next year.
