China is maintaining its lead in the open models space, but the companies with the top models are constantly shifting.
Chinese startup Z.ai’s newly released GLM-4.7 model has claimed the position of the world’s top-performing open-source AI model on the Artificial Analysis Intelligence Index, scoring 68 points and edging past Moonshot AI’s Kimi K2 Thinking model, which scored 67 points. The achievement marks a significant milestone for the Beijing-based company in the increasingly competitive landscape of large language models.
Performance on the Intelligence Index
The Artificial Analysis Intelligence Index evaluates AI models across ten comprehensive benchmarks, including MMLU-Pro, GPQA Diamond, Humanity’s Last Exam, LiveCodeBench, SciCode, AIME 2025, IFBench, AA-LCR, Terminal-Bench Hard, and τ²-Bench Telecom. GLM-4.7’s 68-point score places it ahead of all other open-source models and positions it competitively against several proprietary systems.
In the broader rankings, GLM-4.7 trails only the top-tier proprietary models from major tech companies. Gemini 2.0 Flash Thinking leads the pack at 73 points, followed by Preview High Flash at 73, Gemini 2.0 Flash at 71, Claude 3.7 Sonnet at 70, and o1-High at 70. The model’s performance demonstrates that the gap between open-source and closed-source AI systems continues to narrow.
Technical Advances in GLM-4.7
The new model brings substantial improvements across multiple dimensions compared to its predecessor, GLM-4.6. In coding capabilities, GLM-4.7 achieved 73.8% on SWE-bench (a 5.8 percentage point increase), 66.7% on SWE-bench Multilingual (up 12.9%), and 41% on Terminal Bench 2.0 (a 16.5% improvement). These gains reflect the model’s enhanced ability to handle complex, multilingual coding tasks and terminal-based operations.
Mathematical and reasoning capabilities saw dramatic improvements, with GLM-4.7 scoring 42.8% on the HLE (Humanity’s Last Exam) benchmark with tools enabled—a 12.4 percentage point jump from GLM-4.6’s 30.4% score. The model also demonstrates strong performance on competition-level mathematics, achieving 95.7% on AIME 2025 and 97.1% on HMMT February 2025.
A distinctive feature of GLM-4.7 is its sophisticated thinking architecture, which includes three modes: Interleaved Thinking allows the model to reason before every response and tool call; Preserved Thinking automatically retains reasoning blocks across multi-turn conversations in coding scenarios, reducing information loss; and Turn-level Thinking enables users to toggle reasoning on or off per turn, balancing performance with latency and cost.
Beyond pure performance metrics, Z.ai emphasizes what it calls “Vibe Coding”—the model’s ability to generate cleaner, more modern webpages and better-looking slides with improved layout and sizing. The company argues that true intelligence extends beyond benchmark scores to how seamlessly AI integrates into daily workflows.
Originally known as Zhipu AI, Z.ai is a leading Chinese artificial intelligence company that emerged from Tsinghua University in 2019 as a spin-out startup. The company has established itself as one of China’s “AI Tiger” companies and holds the third-largest share in the large language model market according to International Data Corporation data from 2024.
The company has attracted substantial backing from China’s tech giants, with investors including Alibaba, Tencent, Ant Group, Meituan, Xiaomi, and HongShan. A notable 2023 funding round brought in 2.5 billion yuan (approximately $350 million USD).
Z.ai rebranded internationally in 2025, positioning itself as a global competitor to established AI platforms like ChatGPT and Claude. The company’s GLM model family powers advanced chatbots and assistants, with particular strength in technical domains such as coding, complex reasoning, and content creation. Features like model-native presentation and poster generation distinguish Z.ai’s offerings in the marketplace.
The company’s mission centers on developing artificial general intelligence (AGI) to benefit humanity through machines capable of human-like thinking. Despite its technological advances, Z.ai faces geopolitical challenges—in January 2025, the U.S. Commerce Department added the company to its Entity List over national security concerns, potentially limiting its access to certain American technologies.
Availability and Integration
GLM-4.7 is available through multiple channels. Users can access it via the Z.ai API platform, with comprehensive documentation provided for integration. The model is also available globally through OpenRouter. For those preferring local deployment, model weights are publicly available on HuggingFace and ModelScope, with support for inference frameworks including vLLM and SGLang.
The model integrates with popular coding agents including Claude Code, Kilo Code, Roo Code, and Cline. Z.ai offers a GLM Coding Plan subscription that provides access to GLM-4.7 at what the company claims is one-seventh the price of competing services with three times the usage quota—a pricing strategy reflecting the intense cost competition in China’s AI market.
As the open-source AI landscape continues to evolve rapidly, GLM-4.7’s performance on the Artificial Analysis Intelligence Index demonstrates that Chinese companies remain at the forefront of pushing the boundaries of what publicly available AI models can achieve. Whether Z.ai can maintain this position as competitors like Moonshot AI, DeepSeek, and others release their own updates remains to be seen, but for now, GLM-4.7 is the most capable open model in the world.