Qwen-UI-Agent — Technical Report Website

A concise, application-first website template distilled from the current Qwen-UI-Agent LaTeX draft. Its narrative runs from an interactive real-world capability carousel to benchmark results, broader general and agentic capabilities, playable demos, and release materials, with a complete English / 中文 interface.

Local preview

npm install
npm run dev

Open http://localhost:3000.

Production checks

npm run lint
npm test
npm run build:pages

npm run build:pages creates the static GitHub Pages output in out/. The project is configured for a root Pages site and ships public/.nojekyll. .github/workflows/deploy-pages.yml builds and publishes this output whenever the main branch is pushed.

Editing guide

• Edit all bilingual copy, capability-carousel items, benchmark cards, broader-capability tables, and video links in app/siteContent.ts.
• APPLICATIONS controls the six visual slides in the interactive “what it can do” carousel. Each card intentionally contains only a visual, label, title, and short description.
• Capability visuals support a CSS-drawn mobile scene, a muted looping local video, a three-frame animated browser sequence, and regular image crops.
• MODEL_ORGANIZATIONS maps each benchmark entry to its visible publisher label and local asset in public/brand-logos/.
• GENERAL_CAPABILITY_GROUPS is rendered inside Performance as the “Broader Capabilities” subsection; it is not a standalone page section.
• Desktop performance cards use a three-column layout and collapse responsively on narrower screens.
• The five real-device mobile workflows are hosted directly from public/demos/source/; edit their bilingual titles and instructions in DEMO_VIDEOS.
• The two full-length Computer Use workflows are also hosted from public/demos/source/ as web-optimized 720p H.264 videos. Their complete bilingual task instructions live in DEMO_VIDEOS.
• The Demo explorer uses five domains: real-device mobile, computer use, cross-device GUI use, mobile use with Deep Research, and proactive service. The cross-device domain embeds the two official Bilibili workflows for phone-to-PC receipt processing and mobile-search / PC-summary work. The two proactive workflows are locally hosted portrait videos with bilingual instructions.
• The Deep Research domain embeds the two official Bilibili demos for evidence-based weight-loss verification and World Cup / RedNote retrieval.
• Every DEMO_VIDEOS entry records its original instruction language in instructionSourceLanguage. The opposite-language interface automatically marks the instruction as translated from Chinese or 翻译自英文指令.
• Edit page structure in app/components/ReportPage.tsx.
• Edit the visual system and responsive layout in app/globals.css.
• Replace the Coming soon resource cards with the final technical report, code, and checkpoint URLs at release time.

The primary flow is Capabilities → Performance (including Broader Capabilities) → Demos → Citation.

Every visible navigation, benchmark-domain, demo-category, carousel-control, and status label should provide both English and fully localized Chinese copy. Model and benchmark proper names remain unchanged.

External reference embeds remain explicitly labeled as temporary samples. Result figures that conflict across the current draft are intentionally omitted until the technical-report values are frozen.

The compact benchmark logos are stored locally so the charts do not depend on third-party requests. Most SVGs come from Lobe Icons 1.94.0; Gemini and Anthropic use the supplied reference icons, and the Apodex avatar comes from its official Hugging Face organization. Brand marks remain subject to their respective trademark guidelines and are used only for source identification.