Zhipu AI — operating internationally as Z.ai — just dropped GLM-5V-Turbo, a native multimodal coding model that does one thing exceptionally well: it looks at visual input and writes code from it. Not summaries, not descriptions — actual executable code.
This is the first model in the GLM family built as a native multimodal agent from the ground up. Vision was not bolted on after training. It was baked into the architecture from day one, and it shows.
What GLM-5V-Turbo Actually Does
GLM-5V-Turbo processes screenshots, UI mockups, design files, video, and documents — then outputs runnable code. The primary use cases are design-to-code conversion, GUI automation, and agentic workflows that require visual understanding.
Think of it as a bridge between designers and developers. Hand it a Figma screenshot and it generates the HTML/CSS. Show it a mobile app screen and it produces the component code. Point it at a browser interface and it navigates, clicks, and extracts structured data.
The key differentiator: GLM-5V-Turbo doesn't just see the screen — it understands the interaction model behind what it sees.
Under the Hood: 744B Parameters, 40B Active
GLM-5V-Turbo runs on a Mixture of Experts (MoE) architecture with 744 billion total parameters and 40 billion active per token. The visual encoder is a new component called CogViT, purpose-built for this model.
Key specs at a glance:
| Spec | Value |
|---|---|
| Total Parameters | 744B |
| Active Parameters | 40B per token |
| Context Window | ~203K tokens |
| Max Output | 131K tokens |
| Architecture | MoE with CogViT encoder |
| Quantization | INT8 for inference |
| Training | RL across 30+ task types |
The reinforcement learning approach is notable — Z.ai trained the model jointly across STEM tasks, visual grounding, video understanding, GUI agent workflows, and coding agent tasks. That joint training is why it handles such diverse visual-to-code scenarios.
Benchmark Performance
Z.ai reports a 94.8 score on Design2Code, compared to Claude Opus 4.6's 77.3 on the same benchmark. On GUI agent benchmarks like AndroidWorld and WebVoyager, the model also leads.
A few important caveats here. These are Z.ai's self-reported numbers — independent verification is still pending. And the model reportedly trails Claude in pure-text coding benchmarks, which makes sense given its visual-first optimization.
This is not a general-purpose coding model that happens to see images. It is a vision-first coding model that happens to write decent text-only code.
Pricing: Aggressive
GLM-5V-Turbo costs $1.20 per million input tokens and $4.00 per million output tokens. For context, Claude Opus 4.6 runs $5/$25 and GPT-4o costs $2.50/$10.
Cached input drops to just $0.24/M tokens, making repeated visual analysis workflows remarkably cheap.
| Model | Input/M | Output/M |
|---|---|---|
| GLM-5V-Turbo | $1.20 | $4.00 |
| GPT-4o | $2.50 | $10.00 |
| Claude Opus 4.6 | $5.00 | $25.00 |
For teams running heavy design-to-code pipelines, the cost difference is substantial.
Integration: OpenClaw and Claude Code
GLM-5V-Turbo is optimized to work with OpenClaw, the open-source AI agent framework that recently hit 247K GitHub stars. It also integrates with Claude Code workflows, positioning itself as a specialized visual layer in broader agent stacks.
The API follows the standard OpenAI-compatible chat completions format at https://api.z.ai/api/paas/v4/chat/completions, supporting streaming, function calling, and context caching. SDKs are available for Python, Java, and cURL.
Who Should Care
Frontend teams building design systems will find the Design2Code capabilities immediately useful. QA engineers can leverage the GUI automation for visual regression testing. AI agent builders get a visual perception layer that actually understands UI semantics, not just pixels.
If your workflow involves looking at something and then writing code based on what you see, GLM-5V-Turbo is purpose-built for exactly that.
The Bottom Line
GLM-5V-Turbo is not trying to be the best general coding model — and that is exactly why it is interesting. Z.ai built a specialist, and specialists tend to win in their domain. At $1.20/$4 per million tokens with a 203K context window, the price-to-capability ratio for vision-heavy coding tasks is hard to beat. The self-reported benchmarks are impressive but need independent validation. For now, if you are building anything that bridges the gap between visual design and code, GLM-5V-Turbo deserves a serious look.
