GLM-5V-Turbo generates code from design

# GLM-5V-Turbo: Multimodal Model for Generating Frontend from Design

Chinese company Z.ai has introduced GLM-5V-Turbo — the first multimodal model for coding. It processes images, videos, and text, generating working frontend code as output. Available for free in the web interface at chat.z.ai and via API at $1.2 per 1M input tokens and $4 for output.

The model handles the full cycle: analyzing visual input, planning, and writing code. GLM-5V-Turbo parses design mockups, interface screenshots, and generates complete frontend projects. Combined with Claude Code, it activates URL=Code mode: autonomous site exploration, collecting visual elements, and reconstructing them in code.

Benchmark Results

In Z.ai's internal tests, the model outperforms Claude Opus 4.6 in 9 out of 14 multimodal scenarios. Key metrics:

• Design2Code (mockup to code): 94.8% vs. 77.3%.
• AndroidWorld (GUI agent): 75.7% vs. 62.0%.
• WebVoyager (web navigation): 88.5% vs. 88.0%.

In text-based coding, Claude Opus 4.6 retains the lead: backend, repository analysis, and agent tasks.

| Test | GLM-5V-Turbo | Claude Opus 4.6 |

|------|--------------|------------------|

| Design2Code | 94.8% | 77.3% |

| AndroidWorld | 75.7% | 62.0% |

| WebVoyager | 88.5% | 88.0% |

Technical Architecture

GLM-5V-Turbo includes four improvements:

• Visual encoder CogViT for processing images and videos.
• Reinforcement learning (RL) on 30+ task types: from STEM to GUI agents.
• Agent data generation system for scalable training.
• Extended multimodal toolchain: screenshot capture, bounding boxes, reading web pages with images.

The model comes with 10 built-in skills (Skills):

• Image caption generation.
• OCR and document parsing.
• PDF to presentation conversion.
• Financial report and stock analysis.
• Resume screening.

These tools integrate into the code generation pipeline, boosting agent autonomy.

GLM-5 Lineup and Competition

This release fits into the GLM-5 series:

• GLM-5-Turbo (March): text version for agents.
• GLM-5.1 (late March): coding optimization.
• GLM-5V-Turbo: vision addition.

Direct competition with Claude, Gemini, and Kimi K2.5 in multimodal coding. Z.ai focuses on niche variants of its flagship, strengthening positions in GUI and web development.

For mid/senior developers, the model is handy for UI prototyping: input a Figma mockup or screenshot — get React/Vue/HTML with Tailwind. Test in chat.z.ai for rapid iteration. API integration fits design-to-code workflow automation.

Key Takeaways

• GLM-5V-Turbo leads in Design2Code (94.8%) and GUI tests, surpassing Claude.
• CogViT + RL on 30+ tasks ensure multimodal accuracy.
• 10 built-in Skills expand use beyond coding.
• Free web access + API for production.
• Text coding lags behind Claude; combine models.

— Editorial Team