Beijing Startup Takes Down Silicon Valleys Ai Crown
The Rise of DeepSeek: How a Chinese Startup is Challenging Silicon Valley’s Dominance in AI …
30. July 2025
China Unleashes Crushing Blow: Open-Source Ai Model Leaves Global Rivals In Dust
The Rise of Chinese AI: A New Open-Source Model Crushes Benchmarks
In recent months, the world of artificial intelligence (AI) has witnessed a significant shift in the landscape. China-based startups, backed by giant e-commerce companies like Alibaba, have been making waves with their cutting-edge open-source AI models. The latest addition to this growing list is Ziphu AI’s GLM 4.5 family of AI models, which claims to outperform Anthropic’s highly regarded Opus 4.0 on several benchmarks.
Z.ai, a Beijing-based startup, has unveiled the GLM-4.5 and the GLM-4.5-Air AI models, both based on the Mixture of Experts (MoE) architecture. These models boast unparalleled reasoning, coding, and agentic capabilities, making them a force to be reckoned with in the AI landscape. According to Z.ai, these models are designed to develop a genuinely general model that can excel across various tasks.
The Challenge Facing Traditional GCCs While traditional Generalized Co-Processor (GCC) architectures have been dominating the market for years, they are facing stiff competition from these Chinese startups. As Satvik Paramkusham, an engineer on X, puts it, “It seems like Chinese labs are playing musical chairs at this point.” The reason behind this rapid pace of innovation is the availability of vast resources and expertise in AI research.
A New Era for Open-Source AI The release of Z.ai’s GLM-4.5 model marks a significant milestone in the world of open-source AI. With its unprecedented performance on various benchmarks, it has set the bar high for other models to follow. Moreover, being an open-source model, Z.ai is providing developers with valuable insights into its training process, which is a major plus.
Benchmark Performance The GLM-4.5 model’s performance on benchmark tests is nothing short of impressive. On coding benchmarks, it outperforms models like Claude 4 Opus and OpenAI’s o3. Additionally, it consistently beats DeepSeek-R1 on multiple evaluations, further solidifying its position as a top contender.
User Experiences Users who have had the chance to work with the GLM-4.5 model are overwhelmingly positive about their experience. As one Reddit user put it, “GLM-4.5 is absolutely crushing it for coding – way better than Claude’s recent performance.” Another user on Hacker News reported that they could get the model to consistently use tools and follow instructions in a way that never worked well with Deepseek R1 or Qwen.
Agentic Capabilities The GLM-4.5 model also excels across benchmarks that evaluate its agentic and tool-use capabilities. In tests like BrowseComp, which assesses web browsing capabilities, the model provides correct responses for 26.4% of questions, outperforming Claude 4 Opus (18.8%) and nearing o4-mini-high (28.3%).
The Pelican Benchmark A unique test called the ‘Pelican benchmark’ from Simon Willison, co-creator of Django Web framework, evaluates a model’s practical coding and creative capabilities by asking it to generate an SVG of a pelican riding a bicycle. The GLM-4.5 model impressively passed this test, with one user commenting on how much they liked the design.
Pricing and Availability The GLM-4.5 model is available at an affordable price point of $0.6 per million input tokens and $2.2 per million output tokens, making it a viable option for developers and researchers. The more budget-friendly GLM-4.5-Air variant costs $0.2 per million input tokens and $1.1 per million output tokens.
Reinforcement Learning Infrastructure Z.ai has also open-sourced a reinforcement learning (RL) infrastructure called slime, designed to overcome common RL bottlenecks in complex agentic tasks. Slime’s primary innovations include flexible training architecture, GPU utilization, agent-oriented design, and memory-efficient FP8 format for data generation.
Conclusion The release of Z.ai’s GLM-4.5 model marks a significant turning point in the world of open-source AI. With its unparalleled performance on benchmarks and user-friendly pricing, it has set the bar high for other models to follow. As the landscape continues to evolve, one thing is certain – the future of open-source AI will be shaped by these Chinese startups.