GLM-4.7 Goes Live and Open Source, Delivering a Major Leap in Coding Performance

GLM-4.7, the latest open-source large language model from Zhipu AI, has officially launched, bringing substantial upgrades across coding, reasoning, and agentic execution. The release strengthens GLM’s position as one of the most capable open-source models for real-world software development and tool-driven workflows.

Stronger Coding and Agent Capabilities

GLM-4.7 introduces meaningful improvements in programming performance, particularly in multi-language coding and terminal-based agent workflows. The model now supports a reliable “think before act” mechanism across popular coding frameworks such as Claude Code, TRAE, Kilo Code, Cline, and Roo Code, resulting in greater stability on complex, multi-step tasks.

Key enhancements include:

• Advanced coding performance: More robust code generation and execution across languages and environments.
• Improved front-end aesthetics: Higher-quality generation of webpages, slides, and posters, with better layout, visual balance, and design coherence.
• Stronger tool usage: GLM-4.7 scores 67.5 on the BrowseComp web task benchmark and achieves 87.4 on τ²-Bench for interactive tool use—setting a new open-source state of the art and surpassing Claude Sonnet 4.5.
• Enhanced reasoning: On the HLE (“Humanity’s Last Exam”) benchmark, GLM-4.7 reaches 42.8%, a 41% improvement over GLM-4.6, and outperforms GPT-5.1.
• More natural general intelligence: Conversations are more concise and human-like, with notable gains in writing quality and role-playing immersion. In Code Arena, a blind coding evaluation platform with over one million participants worldwide, GLM-4.7 ranks first among open-source models and first among Chinese models, outperforming GPT-5.2.

Benchmark Results: Competitive with Leading Closed Models

Across mainstream benchmarks, GLM-4.7 aligns closely with Claude Sonnet 4.5 in coding performance:

• SWE-bench-Verified: 73.8% (open-source SOTA)
• LiveCodeBench v6: 84.9% (open-source SOTA, exceeding Claude Sonnet 4.5)
• SWE-bench Multilingual: 66.7% (+12.9%)
• Terminal Bench 2.0: 41% (+16.5%)

Tangible Gains in Real-World Development

1.1 Real Programming Tasks In tests involving 100 real-world coding tasks within the Claude Code environment—covering front-end, back-end, and instruction following—GLM-4.7 demonstrated clear gains over GLM-4.6 in both stability and deliverability. Developers can now organize workflows around true end-to-end task completion, from requirements understanding to production-ready output.

1.2 Controllable Reasoning Evolution GLM-4.7 expands its reasoning system with three modes:

• Interleaved thinking: The model reasons before every response or tool call, improving instruction adherence and code quality.
• Retained thinking: Reasoning blocks persist across multi-turn conversations, boosting cache efficiency and lowering cost for long-running tasks.
• Turn-level thinking control: Developers can toggle reasoning depth per turn—disabling it for simple queries to reduce latency, or enabling it for complex tasks to improve accuracy.

1.3 Complex Task Execution The model shows stronger task decomposition and tech-stack integration, often producing complete, runnable projectsin a single pass, with clear dependency and execution instructions. Demonstrations include fully generated interactive games such as Plants vs. Zombies-style titles and Fruit Ninja-like experiences.

1.4 Front-End Design and Visual Output GLM-4.7 exhibits a deeper understanding of visual code and UI conventions. Default outputs feature improved layout structure, color harmony, and component styling—reducing the need for manual visual fine-tuning. In office productivity scenarios, PPT 16:9 layout accuracy jumped from 52% to 91%, making outputs largely “open-and-use.”

GLM Coding Plan Updated with GLM-4.7

The GLM Coding Plan has been upgraded to include GLM-4.7, offering an optimized balance of performance, speed, and cost:

• Full reasoning support in Claude Code for stable multi-step execution
• Targeted optimizations for Skills, Subagents, and Claude.md workflows
• Built-in visual understanding, search, and web reading for end-to-end coding
• Stronger architecture design and instruction following, reducing hallucinations in long-context scenarios As part of the launch, all paid users receive an experience pass, allowing 3–7 invited users to access a 7-day free trial.

Broad Developer Adoption and Ecosystem Support

Global developer platforms have reported strong results:

• TRAE highlighted GLM-4.7’s improved stability and availability in its China edition.
• Cerebras noted that GLM-4.6 achieved up to 1,000 tokens per second on its hardware, delivering one of the fastest coding experiences available.
• YouWare reported significant gains in front-end design quality, complex feature completion, tool concurrency, and instruction adherence. Additional positive feedback has come from Vercel, OpenRouter, CodeBuddy, and independent developers.

Full-Stack Integration on z.ai and Open Access

GLM-4.7 is now integrated into z.ai with a new Skills module, enabling unified orchestration of multimodal components such as GLM-4.6V, GLM-ASR, and GLM-TTS. This allows developers to build richer, more interactive applications with smoother end-to-end workflows. Access options include:

• API: BigModel.cn
• Online: z.ai, Zhipu Qingyan (app and web)
• Open source:
  • GitHub: zai-org/GLM-4.5
  • Hugging Face: zai-org/GLM-4.7
  • ModelScope: ZhipuAI/GLM-4.7