qwen drops 3.7 plus – the model that reproduces apps on its own

here’s what it means and how they achieved it

the team pointed qwen3.7-plus at apple's native stocks app and told it to reproduce it. no manual instructions, no step-by-step guidance. the model:

• explored the real app autonomously to understand the ui layout and feature logic
• wrote swiftui code from its own interaction records
• pulled in live market data via the longbridge api
• compiled and launched the replica itself
• then ran 10 functional tests – real-time quotes, stock switching, multi-period views, search filtering, detailed stats panel – all passed

how they built this. the underlying system isn't just a coding model. it's what qwen calls a multimodal hybrid agent – a closed loop that looks like this:

see - think - write - act - verify

the model perceives a real screen, reasons about it, writes code, executes gui (graphical user interface) interactions, checks the output, and iterates. gui and cli in a single continuous loop. no context switching, no human in the middle

why this is different from regular coding agents. most coding agents write code and stop. the verification step – does this actually work? does it look right? – gets handed back to a human

qwen3.7-plus closes that gap. it operates the gui to validate what it built, catches failures, and loops back. the human reviews the final output, not every intermediate step

the numbers behind it. on screenspot pro – the benchmark for gui element localization – qwen3.7-plus scores 79.0

• gpt-5.4 sits at 67.4
• gemini-3.1 pro at 68.1

this isn't a marginal win. the model is operationally better at reading and controlling interfaces than the current frontrunners