BEIJING, April 16, 2025 /PRNewswire/ — Z.ai (formerly Zhipu) announces the open-sourcing of its 32B and 9B GLM model series, including base, reasoning, and rumination models, all under the MIT license. These models are now available for free access via the new z.ai website. This platform currently features the 32B base, reasoning, and rumination GLM models and will serve as the primary portal for interacting with and experiencing Z.ai’s latest model releases.
Model Open Source
All models in this open-source release are available under the MIT license. This allows for free commercial use and redistribution, granting developers significant flexibility in usage and development. The release includes models in two sizes, 9B and 32B, encompassing base, reasoning, and rumination models.
The GLM-4-32B-0414 base model, with 32 billion parameters, delivers performance competitive with larger mainstream models globally. Pre-trained on 15 Terabytes (TB) of high-quality data, including extensive synthetic data focused on reasoning, it lays a foundation for future reinforcement learning applications. In the post-training phase, beyond human preference alignment for dialogue scenarios, techniques like rejection sampling and reinforcement learning were employed to specifically enhance performance in instruction following, engineering code generation, and function calls, thereby bolstering the fundamental abilities required for agentic tasks. GLM-4-32B-0414 excels in engineering code generation, Artifacts creation, function calls, search-based question answering, and report writing, with some benchmark scores approaching or exceeding those of larger models like GPT-4o and DeepSeek-V3-0324 (671B).
GLM-4-32B-0414 has further improved code generation capabilities, capable of handling and generating more complex single-file code. Z.ai’s dialogue mode includes a preview feature, supporting visualization of generated HTML and SVG, facilitating user evaluation of results and iterative optimization.
GLM-Z1-32B-0414 is a reasoning model featuring advanced reasoning capabilities. Built upon GLM-4-32B-0414, it utilizes cold-start and extended reinforcement learning strategies and has undergone deep optimization training focused on key areas like mathematics, coding, and logic. Compared to the base model, GLM-Z1-32B-0414 shows significant enhancements in mathematical reasoning and complex problem-solving capabilities. Furthermore, the training integrated general reinforcement learning techniques leveraging pairwise critic rewards, effectively boosting the model’s overall general-purpose abilities.
In certain tasks, GLM-Z1-32B-0414, with its 32 billion parameters, matches the performance of DeepSeek-R1, which has 671 billion parameters. Through evaluations in benchmark tests such as AIME 24/25, LiveCodeBench, and GPQA, GLM-Z1-32B-0414 has demonstrated strong mathematical reasoning abilities, enabling it to address a broader range of complex tasks.
The GLM-Z1-9B-0414, a 9B model, delivers impressive performance for its size. Trained using the same advanced techniques applied to larger models, this smaller variant demonstrates excellent capabilities in mathematical reasoning and general tasks despite its reduced parameter count, ranking it among the leading open-source models in its size class. Particularly for resource-constrained environments, it strikes an effective balance between capability and efficiency, offering a strong option for users requiring lightweight deployment.
The rumination model GLM-Z1-Rumination-32B-0414 represents Z.ai’s next step in exploring the future form of AGI.
Unlike standard reasoning models, the Rumination model tackles highly open-ended and complex problems via a more extended, multi-step deliberation process. A key innovation is its ability to integrate search tools during this process to address complex tasks, guided by various rule-based reward mechanisms that steer and enhance end-to-end reinforcement learning. The model supports a full research cycle—autonomously formulating questions, searching for information, constructing analyses, and completing tasks—significantly boosting its proficiency in research-oriented writing and complex information retrieval.
Z.ai Launched
The new z.ai website, serving as the primary portal for interacting with Z.ai’s latest models, is now live!
Z.ai currently features three open-source models for interactive use:
- GLM-4-32B (Base Model): Features powerful code generation capabilities, supports the new Artifacts feature, and enables an interactive development experience.
- Z1-32B (Reasoning Model): Delivers superb reasoning performance, allowing users to experience inference speeds up to 200 Tokens/second online.
- Z1-Rumination-32B (Rumination Model): Allows users to explore the powerful “Deep Research” capabilities, ideal for conducting in-depth investigations.
About Z.ai
Founded in 2019, Beijing Zhipu Huazhang Technology Co., Ltd. (Z.ai, formerly Zhipu) is a leading Chinese AI company focused on developing next-generation cognitive models. Since launching the GLM pre-training framework in 2020, Z.ai has released several industry-leading models, including GLM-130B and ChatGLM, with over 30 million global downloads. Its product suite includes QingYan, CodeGeeX, CogVLM, and CogView, as well as an innovative Model-as-a-Service (MaaS) platform for developers and enterprises. Recognized with numerous industry awards, Z.ai is driving progress towards the era of artificial general intelligence.
For more information, visit zhipuai.cn/en/.