Alibaba's Qwen3.7-Plus Beats GPT-5.4 at GUI Automation by 12 Points

Most multimodal models follow a predictable pattern: bolt a vision encoder onto a language model, let it describe images, and call it done. The output is still text. The model is still passive. You still need a separate pipeline to turn what it sees into something executable. Qwen3.7-Plus is built differently , and the gap matters more than it might sound.

One model to see, think, code, and act

Qwen3.7-Plus is a multimodal agent model that unifies vision and language into a single, versatile agent foundation. Building on Qwen3.7's strong text backbone, it delivers a comprehensive upgrade in vision-language capabilities while retaining full agentic strength in coding, tool use, and productivity workflows. The key word is agent. This is not a vision-augmented chatbot , it is a model designed to close the loop between perception and execution.

What sets Qwen3.7-Plus apart is its ability to operate as a multimodal interactive hybrid agent. It perceives real-world scenes, reads screens and operates GUIs, writes code from visual references, navigates mobile apps end-to-end, and answers visual questions grounded in web knowledge , seamlessly blending GUI and CLI interactions within a single agent loop.

The five agentic capabilities layered on top of visual understanding are: deep reasoning, self-programming (the model writes and revises its own code), tool invocation (calling external APIs), verification and testing (running outputs and checking results), and autonomous iteration (looping until the task is done). These aren't spec-sheet features , they're the components of an agent that can take a task from screenshot to shipped result with no human in the loop.

The architecture that makes it work

Unlike multimodal models that bolt vision on top of a text-first architecture, Qwen3.7-Plus was trained with early fusion: vision and language tokens are processed together from the first layer, not integrated at a late stage. Alibaba reports the model was trained on trillions of multimodal tokens with this approach. The practical effect: the model does not just describe images , it reasons about them with the same chain-of-thought depth as the text-only Max.

The 1M-token context window, carried over from the rest of the Qwen3.7 family, means these multi-step workflows don't collapse under the weight of their own context. Long traces from 1,000+ tool calls stay coherent. Prior code and intermediate outputs remain accessible. This is the architectural detail that separates a capable demo from a production-grade agent.

It generalizes across agent scaffolds, performing consistently whether deployed through Claude Code, OpenClaw, Qwen Code, or other frameworks. For teams already using a specific agent harness, that means a model string swap is often all it takes.

The benchmark case: what the numbers actually measure

The most informative benchmarks for Qwen3.7-Plus are not the standard language leaderboards , they're the ones that measure what the model does with visual input, not just whether it can describe one.

• ScreenSpot Pro (GUI grounding): Qwen3.7-Plus scores 79.0% , above GPT-5.4 (67.4%), Gemini-3.1 Pro (68.1%), and Qwen3.6-Plus (68.2%). ScreenSpot Pro measures the model's ability to look at a screenshot of a software application and produce the exact pixel coordinates of the UI element to interact with , the bottleneck capability for everything called "computer use" or "browser automation."
• AndroidWorld (mobile task completion): Qwen3.7-Plus scores 81.0% , above Gemini-3.1 Pro (70.7%) and Qwen3.6-Plus (67.2%).
• Terminal-Bench 2.0 (agentic terminal execution): It scored 70.3 at launch , the highest of all evaluated models including Claude Opus-4.6 Max (65.4%) and DeepSeek-V4-Pro Max (67.9%).
• Deep-Planning (multi-step planning under uncertainty): 62.3% , best-in-class across evaluated frontier models.
• OmniDocBench 1.5 (document understanding): 91.4% , the highest of all evaluated models, above Gemini-3.1 Pro (90.0%) and Claude Opus-4.6 Max (86.6%).
• MathVision (visual math reasoning): 90.3%, second only to GPT-5.4 (91.0%).
• SimpleVQA (vision + live search): 81.7%, above GPT-5.4 (69.4%) and Gemini-3.1 Pro (76.9%).

Qwen3.7-Plus scored 79.0 on ScreenSpot Pro. A 12-point lead over GPT-5.4 on GUI grounding is not incremental , it places Qwen3.7-Plus at the top of the non-Western multimodal field for this specific capability.

What 11 hours of autonomous development looks like in practice

Benchmarks describe capability. Alibaba's internal demo makes it concrete. The task: build a complete English vocabulary learning app from scratch, end to end, with no human intervention. The agent operated continuously, generating more than 10,000 lines of code and triggering over 1,000 agent calls, completing every stage of a real software development lifecycle: requirement document generation, automated coding, installation and deployment, test case creation, GUI-based automated testing, multi-scenario parallelized testing, automatic product documentation updates, and autonomous version iteration.

A second demo is arguably more precise as a technical proof point. The task: reproduce the native macOS Stocks application using only autonomous interaction with the original app as input. The agent studied the UI layout, wrote SwiftUI source code from those interaction records, integrated a live market data API, compiled and launched the reproduced app, then ran 10 functional verification tests autonomously. All 10 passed. Dark theme, split-view layout, real-time data, full interactivity , faithfully reproduced.

This is what ScreenSpot Pro 79.0% and AndroidWorld 81.0% mean in practice.

How GUI grounding actually works under the hood

Qwen3.7-Plus produces structured action plans , "click at (x=487, y=232)" or "type [email protected]" , that an external runtime executes through tools like Playwright, Selenium, PyAutoGUI, or Appium. The model handles perception and decision-making; the runtime handles physical interaction. This is the same architecture used by Claude Computer Use and OpenAI Operator.

The generate-run-check loop is what separates a vision-to-code model from a useful vision-to-code agent. On benchmarks like OSWorld-Verified and AndroidWorld, this loop is precisely what is being evaluated , not just "did it write code that looks right" but "did it produce something that works."

Speed, cost, and what's closed

On agent tasks, end-to-end throughput for Qwen3.7-Plus reached 147.5 tokens/s, while Qwen3.6-Plus reached only 41.5 t/s , a 3.55x improvement. That's not just a benchmark win; it means long autonomous runs complete faster and cost less per task.

Pricing sits at $0.40/$1.60 per million input/output tokens, with $0.08 for cached input reads , roughly 6x cheaper than Max on input, 4.7x cheaper on output. For budget-sensitive agent pipelines that need vision, that gap is significant.

There is one important caveat: Qwen3.7-Plus ships proprietary and API-only, breaking the lab's open-weight habit. You cannot run it on your own infrastructure, inspect its behavior at the weight level, or verify training data claims independently. For builders evaluating Chinese AI infrastructure, data routing goes through Alibaba Cloud, and the model's behavior is not independently auditable. That's a real consideration for teams with data sovereignty requirements.

Who should use it and for what

The use cases break down cleanly by what the model uniquely enables:

• Frontend and design-to-code teams: Feed a Figma export, a reference screenshot, or a video walkthrough and get functional HTML, CSS, React, or SwiftUI , without a manual transcription step.
• Agent builders: Terminal-Bench 70.3% and Deep-Planning 62.3% are both best-in-class. If you're building agents that need to sustain long-horizon, multi-step execution across real environments, this is the strongest foundation currently available.
• Enterprise document pipelines: OmniDocBench 91.4% means it handles invoices, contracts, research papers, and financial reports with best-in-class accuracy , and its multimodal search integration lets it supplement extraction with live web knowledge.
• GUI and mobile automation teams: AndroidWorld 81.0% and ScreenSpot Pro 79.0% translate directly to production-realistic reliability for navigating real mobile apps and desktop interfaces.
• Multilingual products: PolyMATH 84.0% (highest of all evaluated models) and WMT24++ 84.6% mean the model's reasoning and agent capabilities hold up across non-English contexts.

The bigger picture

The standard assumption about multimodal models has been that vision is an add-on , useful for specific tasks, but not architecturally central. Qwen3.7-Plus challenges that directly. Whoever wins the GUI automation surface will likely define the default architecture for agent products in 2027.

The deeper shift is this: vision, combined with tool use and long-context coherence, changes what an agent can be assigned to do. A model that can see a UI element, decide to click it, click it, check the result, and revise , all within the same context window , eliminates the human-in-the-loop steps that currently gate most automation workflows. Qwen3.7-Plus is among the first models that makes this concrete at a price point that doesn't require a frontier budget. It is available on Alibaba Cloud Model Studio and Qwen Chat, with API access under alibaba/qwen3.7-plus.