Zhipu AI Open-Sources 32B/9B GLM Series Models, Launches Z.ai for Free Access — A Game-Changer in the AI World!
Zhipu AI has open-sourced its 32B/9B series GLM models, covering base, reasoning, and deep-thought models, all under the permissive MIT license. These models are now freely accessible via the brand-new platform Z.ai and have also been integrated into Zhipu’s MaaS (Model-as-a-Service) platform at bigmodel.cn. Among them, the reasoning model GLM-Z1-32B-0414 delivers performance comparable to top-tier models like DeepSeek-R1, with real-world inference speeds reaching an impressive 200 tokens/second—currently the fastest among domestic commercial models. Even more striking, its pricing is just 1/30 of DeepSeek-R1’s.
Zhipu AI has launched the new domain Z.ai, which currently hosts three types of GLM models: base, reasoning, and deep-thought models. This platform will serve as the primary interactive experience hub for Zhipu’s latest models moving forward.
Key Highlights of the Open-Sourced Models
All models released under this initiative are governed by the MIT license, allowing free use for commercial purposes and unrestricted distribution. This provides developers with unparalleled freedom for usage and further development. The open-sourced models include two sizes—9B and 32B—with variations for base, reasoning, and deep-thought tasks:
- 1. Base Model: GLM-4-32B-0414
- • 32 billion parameters, delivering performance on par with mainstream models both domestically and internationally, even those with significantly larger parameter counts.
- • Pre-trained on 15T of high-quality data, including extensive synthetic reasoning datasets, laying a strong foundation for reinforcement learning extensions.
- • Enhanced through post-training techniques such as human preference alignment, rejection sampling, and reinforcement learning, improving capabilities in instruction-following, code generation, function calling, and more.
- • Excels in tasks like engineering code, artifact generation, function calls, search-based Q&A, and report writing. Some benchmark metrics rival or surpass larger models like GPT-4o and DeepSeek-V3-0324 (671B).
- • Advanced code generation capabilities allow handling and generating more complex single-file code. On Z.ai, users can preview generated HTML and SVG outputs for iterative optimization.
- 2. Reasoning Model: GLM-Z1-32B-0414
- • Built upon GLM-4-32B-0414, this model leverages cold-start and extended reinforcement learning strategies, with a focus on math, coding, and logical reasoning tasks.
- • Demonstrates significant improvements in mathematical reasoning and problem-solving abilities, matching the performance of DeepSeek-R1 (671B) in certain tasks despite having only 32B parameters.
- • Evaluated on benchmarks like AIME 24/25, LiveCodeBench, and GPQA, showcasing robust reasoning capabilities for tackling complex challenges.
- 3. Smaller but Mighty: GLM-Z1-9B-0414
- • Despite its compact size of 9 billion parameters, this model delivers outstanding performance in mathematical reasoning and general tasks, ranking among the top open-source models of its size.
- • Ideal for resource-constrained environments, offering an excellent balance between efficiency and effectiveness, making it perfect for lightweight deployments.
- 4. Deep-Thought Model: GLM-Z1-Rumination-32B-0414
- • Represents Zhipu AI’s next step toward exploring AGI (Artificial General Intelligence).
- • Solves highly open-ended and complex problems through multi-step deep thinking, integrating search tools and rule-based reward mechanisms to guide end-to-end reinforcement learning.
- • Supports a complete research loop: “posing questions → searching for information → analyzing data → completing tasks.”
- • Particularly excels in research-oriented writing and complex retrieval tasks.
Performance Breakthroughs
- • Inference Speed: The reasoning model GLM-Z1-AirX achieves an industry-leading speed of 200 tokens/second, 8x faster than standard models.
- • Cost Efficiency: The GLM-Z1-Air version offers pricing at just 1/30 of DeepSeek-R1, ideal for high-frequency API calls.
- • Free Tier: The GLM-Z1-Flash version is completely free, lowering barriers to entry for developers and businesses.
Introducing Z.ai
The newly launched Z.ai serves as the gateway to interact with Zhipu AI’s cutting-edge models. Currently available models include:
- 1. GLM-4-32B (Base Model): Powerful code generation with interactive Artifacts functionality.
- 2. Z1-32B (Reasoning Model): Ultra-fast inference with up to 200 tokens/second.
- 3. Z1-Rumination-32B (Deep-Thought Model): Experience advanced deep-research capabilities for complex analysis.
Why This Matters
By combining open-source accessibility, industry-leading performance, and cost-effective solutions, Zhipu AI is setting a new standard in the AI landscape. Whether you’re a developer, researcher, or enterprise user, these models provide unparalleled flexibility and power to drive innovation.
Visit Z.ai today to explore the future of AI!
Reproduction without permission is prohibited：AI LAB » Zhipu AI Releases Powerful 9B/32B Series Models as Open Source, Launches Z.ai to Shake Up the AI World!
