Zhipu AI has released GLM-4.5, a new family of open-source language models that rivals U.S. and European systems in reasoning, coding, and multimodal tasks.
The launch raises the stakes in the global AI race, with China’s billion-dollar startup now openly challenging OpenAI, Anthropic, and Google.
Key Takeaways:
- Zhipu AI unveils GLM-4.5 models with coding and agent abilities.
- Multimodal GLM-4.5V adds video, image, and website reconstruction.
- Benchmarks show high efficiency with fewer parameters than rivals.
- Backed by Alibaba, Tencent, Xiaomi — firm valued over $5 billion.
- Raises global AI rivalry as IPO prep and U.S. restrictions loom.
Zhipu AI’s GLM-4.5 family is an open-source set of Chinese AI models optimized for reasoning, programming, and multimodal tasks. With fewer parameters than Western rivals, GLM-4.5 matches or outperforms models like Deepseek-R1 and even Claude Opus in benchmarks, signaling China’s accelerating push to close the AI gap.
China’s Zhipu AI Enters the Global AI Arms Race
Beijing-based Zhipu AI has officially launched GLM-4.5, a suite of open-source language models designed to take on leading Western competitors like OpenAI’s GPT-4 and Anthropic’s Claude.
The release includes three versions:
- GLM-4.5 (standard)
- GLM-4.5-Air (lightweight)
- GLM-4.5V (multimodal with vision and video)
Users can try them for free through chat.z.ai, in a ChatGPT-style interface.
Features and Capabilities
The standout GLM-4.5V offers image and video analysis, website reconstruction from screenshots, and on-screen operations for autonomous agents.
Practical demos highlight real-world use: generating mini-games, building physics simulations, and producing presentation slides with autonomous search.
Performance Benchmarks
Despite fewer parameters, GLM-4.5 ranks near the top in independent benchmarks:
- 70.1% on TAU-Bench agent tasks
- 91.0% on AIME 24 math problems
- 64.2% on SWE-Bench Verified software engineering tasks
In web navigation, GLM-4.5 even surpassed Claude Opus 4, scoring 26.4% vs 18.8%.
“This efficiency challenges the assumption that bigger always means better in AI”.
Architecture: Depth Over Width
Unlike Western peers, Zhipu favors deeper neural networks over wider ones. GLM-4.5 uses a mixture-of-experts model with 355B parameters (32B active). The lighter Air version runs at 106B (12B active).
Training spanned 23 trillion tokens, progressing from general data to specialized reasoning and code.
Investor Backing and IPO Path
Founded in 2019 by Tsinghua University professors, Zhipu has grown into a 5B+ USD company with 800+ employees, most in R&D.
Heavyweight investors include Alibaba, Tencent, Xiaomi, and international backers like Saudi Aramco’s Prosperity7 Ventures. An IPO is reportedly on the horizon.
Human Reaction
When Zhipu released early demos, Chinese developers flocked online. One user joked on Weibo: “It feels like ChatGPT finally speaks Mandarin natively — but it also builds my whole website!”
Why It Matters
The GLM-4.5 release illustrates how China is rapidly closing the AI gap with U.S. firms. While OpenAI and Anthropic dominate headlines, Zhipu’s efficiency-first approach shows another path to high-performance AI that doesn’t rely solely on parameter size.
Impact
For everyday users, this could mean faster, cheaper, and more versatile AI tools in the coming year. Whether you’re a student, developer, or small business owner, the availability of open-source systems like GLM-4.5 widens access — though censorship rules may limit outputs.
Numbers to Watch
- $5 billion+ → Zhipu AI’s current valuation
- 355B parameters → GLM-4.5 size (32B active at inference)
- 91.0% → Score on AIME 24 math benchmark
What’s Next
IPO preparations in China by 2025.
Possible new multimodal features, including voice.
Rising geopolitical tension as U.S. eyes tighter AI export controls.
Developers worldwide testing GLM-4.5 as an open-source alternative.
Conclusion
Zhipu AI’s GLM-4.5 launch is more than just another model drop. It signals a serious Chinese push to rival Western AI giants, combining strong academic roots, heavy corporate backing, and parameter-efficient design.
As U.S.–China AI rivalry intensifies, the question isn’t just who builds the smartest models — but who controls the values, censorship rules, and global adoption that shape AI’s future.