Z.ai Unleashes GLM-4.7: The New Open-Source Standard for Coding and Reasoning
GLM-4.7 is here!
— Z.ai (@Zai_org) December 22, 2025
GLM-4.7 surpasses GLM-4.6 with substantial improvements in coding, complex reasoning, and tool usage, setting new open-source SOTA standards. It also boosts performance in chat, creative writing, and role-play scenarios.
Default Model for Coding Plan:… pic.twitter.com/3vDzwof7A8
On December 22, 2025, the AI landscape shifted as Zai.org (formerly Zhipu AI) announced the release of GLM-4.7. This next-generation large language model (LLM) represents a massive leap over its predecessor, GLM-4.6, specifically targeting the most demanding frontiers of artificial intelligence: autonomous software engineering, complex reasoning, and agentic tool usage.
By claiming state-of-the-art (SOTA) benchmarks and even outperforming proprietary giants like GPT-5.1 in specific categories, GLM-4.7 is positioning itself as the premier choice for developers who demand high-performance, open-source intelligence.
Model Background: From Zhipu to Z.ai
Z.ai has long been a powerhouse in the Chinese AI ecosystem, gaining global recognition for its "General Language Model" (GLM) series. The transition from GLM-4.6 to 4.7 is more than an incremental update; it is a refined 358-billion parameter powerhouse.
The model's development reflects Z.ai's core mission: achieving Artificial General Intelligence (AGI) through an open-source philosophy. By making these weights available on platforms like Hugging Face, Z.ai is challenging the dominance of closed-model providers, offering "frontier-level" performance without the proprietary lock-in.
Key Improvements: Coding, Vibes, and Tools
GLM-4.7 introduces several critical enhancements that make it uniquely suited for "Real Development Workflows":
- Software Engineering Excellence: The model saw a +5.8% increase on the SWE-bench Verified leaderboard compared to GLM-4.6, proving its ability to resolve real-world GitHub issues autonomously.
- "Vibe Coding" & Creative UI: Beyond logic, GLM-4.7 excels at front-end generation. Whether creating high-contrast dark mode websites or complex voxel art (such as a pagoda garden), the model translates aesthetic intent into functional code with unprecedented accuracy.
- Agentic Tool Usage: With a score of 87.4% on τ²-Bench, GLM-4.7 demonstrates a superior ability to navigate tool-based environments, making it a highly reliable engine for AI agents.
- Deep Reasoning: On the Hard LLM Benchmark (HLE), when equipped with tools, the model reached a score of 42.8%, surpassing many of its contemporaries in solving complex, multi-step logical problems.
Benchmark Results: GLM-4.7 vs. The Giants
The performance of GLM-4.7 puts it in direct competition with the industry's most advanced proprietary models, including Claude Sonnet 4.5 and GPT-5.1.
| Benchmark | GLM-4.7 | GLM-4.6 | Claude Sonnet 4.5 | GPT-5.1 High |
|---|---|---|---|---|
| SWE-bench Verified | 73.8% | 68.0% | 77.2% | 76.3% |
| HLE (w/ Tools) | 42.8% | 30.4% | 32.0% | 42.7% |
| τ²-Bench (Tool Use) | 87.4% | 75.2% | 87.2% | 82.7% |
| Terminal Bench 2.0 | 41.0% | 24.5% | 42.8% | 47.6% |
Note: GLM-4.7 notably matches or exceeds GPT-5.1 in Tool Usage (τ²-Bench) and Reasoning (HLE).
Technical Innovation: The "Thinking" Architecture
What sets GLM-4.7 apart is its sophisticated reasoning architecture, which allows it to "pause and think" before and during a response:
- Interleaved Thinking: The model performs pre-response reasoning, mapping out a logic chain before generating the final answer to ensure consistency.
- Preserved Thinking: For agentic workflows, the model retains its internal reasoning logic across multiple turns, preventing the "forgetting" common in long-horizon tasks.
- Turn-level Thinking: A toggleable feature that allows developers to control latency. Users can choose between a fast "instinctual" response or a slower, "deep-thinking" response for complex problems.
The model also natively supports inference frameworks like vLLM and SGLang, ensuring it can be integrated into existing high-speed production environments.
Showcase: Real-World Applications
Z.ai has released several demos showcasing GLM-4.7’s capabilities:
- Dynamic Presentations: Generating a full set of Zootopia-themed intro slides with consistent layouts.
- Creative Engineering: Creating 3D Voxel Art of a pagoda garden from a simple text prompt.
- Front-end Development: Building a functional, themed dashboard with complex UI components in seconds.
Access and Deployment
Z.ai has ensured that GLM-4.7 is accessible to everyone, from hobbyists to enterprise developers:
- Try it Now: Available at chat.z.ai. It is currently the default model for users on the "Coding Plan."
- API: Developers can access the model via Z.ai’s API or through OpenRouter.
- Open Source: The model weights (including an optimized FP8 variant) are available on Hugging Face.
- Cost Efficiency: The GLM Coding Plan offers roughly three times the quota of Claude Sonnet at nearly 1/7th of the cost, making frontier-level coding accessible to a wider audience.
The Future of Open-Source AGI
The release of GLM-4.7 is a milestone for the open-source community. By rivaling the reasoning and coding capabilities of the world’s most expensive proprietary models, Z.ai is proving that the gap between open and closed AI is closing faster than expected.
As rumors of a Z.ai IPO continue to circulate, the success of GLM-4.7 reinforces their position as a global leader in the race toward AGI. For developers and enterprises looking to build the next generation of AI agents, GLM-4.7 isn't just a new model—it’s a new benchmark for what's possible.