Alibaba shifts Qwen toward full agent workflows
Alibaba’s Qwen team has released Qwen3.7-Plus, a new multimodal model that combines visual understanding with classic agent capabilities such as coding and tool use. The company describes it as a multimodal interactive hybrid agent, and the positioning is notable: this is not framed as a chatbot with image input, but as a system meant to perceive interfaces and act inside them.
According to the supplied source text, Qwen3.7-Plus is designed to recognize real-world scenes, read screen content, operate graphical user interfaces, write code from visual templates, and navigate mobile apps end to end. The operating model matters. UI clicks and command-line instructions run inside the same agent loop, which suggests Alibaba is aiming at a more unified form of automation rather than separate models for perception, planning, and execution.
Long-running tasks are central to the pitch
Alibaba’s showcase examples focus on autonomy over extended workflows. In one demonstration, a hybrid agent system built an English vocabulary learning app over more than eleven hours. The source says the run produced more than 10,000 lines of code across more than 1,000 agent calls.
The reported process covered requirements documentation, automated code generation, dependency installation, test-case creation, GUI-based testing, parallel test scenarios, and version management. Those details are important because they push the story beyond a one-shot coding demo. Alibaba is arguing that the model can persist through a multi-stage software project and keep operating across tools and interfaces without repeated human hand-holding.
A second demonstration moved from software generation to software imitation. Alibaba says the agent recreated Apple’s native macOS Stocks app by parsing the interface, generating SwiftUI code, connecting an external real-time stock data API, compiling the result, and running ten functional tests on its own. If that performance generalizes, the model’s value may lie less in answering prompts and more in compressing the time between seeing a working interface and reproducing it in code.
Browser and cloud operations broaden the scope
The third use case extends the model into browser-based operations. Through a sidebar extension called Qwen for Chrome, the system can switch into agent mode with user permission and perform cloud console tasks. The source text cites one example in which the model purchased the cheapest available virtual server instance, including setup of image, storage, and security-group options.
Alibaba also says the model handled follow-up scaling and maintenance tasks. That matters because it moves the pitch from isolated task completion to lifecycle management. A model that can create, test, configure, and later maintain a service crosses into territory that enterprises have typically reserved for combinations of engineers, scripts, and workflow tools.
Strong GUI performance, weaker pure reasoning
The benchmark picture in the supplied material is mixed. Alibaba’s published results reportedly show Qwen3.7-Plus performing especially well on graphical interface tasks. On AndroidWorld and ScreenSpot Pro, the model is described as sitting well ahead of GPT-5.4 (xhigh). That gives Alibaba a concrete angle in a crowded market: if interface manipulation becomes a major AI battleground, Qwen wants to compete on execution, not only conversation.
At the same time, the source text says the system falls short in pure logic benchmarks. That caveat is significant. It suggests that Qwen3.7-Plus may be more useful when the environment itself provides structure, visual anchors, and action affordances than when the model must solve abstract reasoning tasks without that context.
In practical terms, the model’s strengths appear to be grounded in seeing software and acting inside it. That is a narrower but commercially relevant definition of intelligence, especially for enterprise automation, testing, customer operations, and software prototyping.
Why the release matters
Qwen3.7-Plus is also positioned as a proprietary but comparatively inexpensive option through Alibaba Cloud. Price and deployment path matter because agentic systems can become costly quickly when they run long sessions, execute many calls, and interact with external tools. If Alibaba can keep operating costs down while delivering strong interface performance, it may find a receptive market among developers and businesses that want automation without frontier-model pricing.
The broader significance is that Qwen3.7-Plus reflects a shift in how AI vendors are defining progress. Instead of focusing only on benchmark scores or chat quality, Alibaba is emphasizing whether a model can observe an interface, make decisions, call tools, write code, and stay on task for hours. That does not settle the harder questions around reliability, oversight, and failure handling. But it does show where the competition is heading: toward AI systems judged by what they can complete, not just what they can say.
This article is based on reporting by The Decoder. Read the original article.
Originally published on the-decoder.com