How HarmonyOS Partners Cut Cold Starts to 1s & Boost GPU 17x at HDC2026

HDC2026 Developer Scenario Technology Co-construction Forum Overview

On June 13, 2026, HarmonyOS ecosystem partners gathered at Songshan Lake for the "Huawei Developer Conference HDC2026 — Developer Scenario Technology Co-construction Forum." The forum invited application partners and Huawei experts to share practical cases, dissect core technologies, and exchange co-construction experiences around AI-assisted R&D and media scenario optimization.

Opening Keynote: Co-construction Achievements

Xu Jiazhe, Secretary of the OpenHarmony Project Group Technical Steering Committee, presented "Join Hands to Co-construct, Share Prosperity in HarmonyOS Ecosystem." He reported the past year's results: partners shifted from merely using to actively co-constructing, creating 160+ innovative scene capabilities , enhancing the innovative competitiveness of 1,600+ applications , and achieving a 90x increase in partner usage rate . These efforts significantly improved HarmonyOS app user experience and accelerated development efficiency, pushing the ecosystem into a flywheel-accelerated scaling era.

AI-Assisted R&D Co-construction: Full-Chain Efficiency Gains

Amap AGenUI: Generative UI Framework

Teng Tengjixiang, full-stack R&D engineer at Amap, presented "Embracing Intelligence: Amap AGenUI Opens a New Era of Generative UI on HarmonyOS." Leveraging travel scenarios, Amap built the AGenUI framework that evolves AI applications from simple "text-based interaction" to efficient, intuitive "generative UI interaction." On HarmonyOS, the framework delivers 20% rendering performance improvement and 18% memory reduction , demonstrating the new possibilities of generative UI for HarmonyOS app interaction — shifting from "one-size-fits-all" to "personalized for each user."

WPS "Diting" AI Flame Graph Analysis

Fu Hao, WPS PC client R&D engineer, presented "WPS: Diting x AI Helps You Understand Flame Graphs for Automatic Performance Bottleneck Localization." Previously, analyzing a single performance trace required senior engineers to spend hours manually; now AI automatically parses flame graphs and precisely pinpoints bottlenecks. Results: single test case analysis time reduced from 5 minutes to 1 minute ; analysis of thousands of samples compressed from 66 hours to 13 hours ; each performance specialization round saves over 130 man-hours . With AI assistance, every developer can become a performance tuning expert.

Hardcore Performance Optimization: Startup, Network, GPU

Tencent Video: Cold Start from 2.x Seconds to 1 Second

Wang Wenshuai, Tencent Video infrastructure lead, shared "Challenging Long-Video App Cold Start Limits: The 2.xs → 1s Performance Leap." Four key techniques were employed:

Cross-language zero-copy calls to eliminate data copying overhead.

Network pre-connection pre-warming to establish connections before they are needed.

Module lazy loading to defer non-critical initialization.

Business unitization refactoring to isolate and optimize startup-critical paths.

These measures collectively reduced cold start time from over two seconds to 1 second , achieving a "second-open" experience.

Meituan: QUIC Long-Connection Optimization

Hu Zejiang, Meituan network expert, presented "Meituan Client Long-Connection QUIC Optimization." Targeting food delivery and flash purchase scenarios that suffer from slow loading and message loss under weak networks, Meituan optimized client long connections based on the QUIC protocol. Outcomes: weak-network page load latency improved by 27% , network switching became imperceptible to users , enabling more timely orders and smoother shopping.

Yixin Technology: SimpleGPULayer High-Performance Image Processing Base

Huang Wei, Yixin Technology technical lead, shared "Wukong Image x SimpleGPULayer High-Performance Image Processing Base Evolution." Co-developed with the Huawei HarmonyOS team, SimpleGPULayer provides a unified abstraction layer that encapsulates Vulkan low-level capabilities. Developers need only 3 lines of code to invoke full GPU hardware acceleration. The framework includes 52 high-performance GPU image filters covering 90% of mainstream professional image processing needs, and 28 AI inference compute operators . For high-frequency complex scenarios, it achieves up to 1713% (17x) speedup .

Partner-Contributed Open Source Scene Capabilities

Huajiao Live: HJPusher & HJPlayer SDK

Gao Shang, Huajiao Live multimedia technology expert, presented "One Push One Play, Smart Connection to HarmonyOS: Huajiao Live HJPusher & HJPlayer SDK Practice." The SDK covers the full live streaming chain — anchor-side pushing and audience-side playback — providing complete capabilities from capture, encoding, pushing to pulling, playback, and rendering. A plugin-based architecture makes the pipeline composable and capabilities reusable, lowering secondary development barriers. Optimizations target first-frame latency, anti-jitter, and low latency. The SDK fully leverages HarmonyOS multimedia and AI Kits to release system capabilities, delivering a mature live streaming solution for the ecosystem.

Qinbaobao: ImageLoader Image Loading Optimization

Yu Xin, Qinbaobao HarmonyOS development lead, presented "From 'Can Load' to 'Instantly Visible': ImageLoader Image Loading Optimization Practice." Addressing common HarmonyOS app pain points — white blocks, flickering, wrong images, slow loading — Qinbaobao developed the unified ImageLoader component. A single API covers network images, local files, video covers, and other data sources. Multiple optimization techniques make image loading smoother and more stable under weak networks.

Yuandi Qiu: Near-Field Media File Sharing & KML Parsing

Lin Zhibiao, Yuandi Qiu founding thread founder, shared "Media File Near-Field Sharing" — two open source contributions:

Tap-to-share based on HarmonyOS touch capability : enables one-tap file sharing, receiving, and auto-import. Frontline workers complete template distribution within 10 seconds without understanding paths, compression packages, or other operations.

High-performance KML parsing engine optimized for GIS scenarios : parses 10,000+ geographic features in milliseconds. Integration cost reduced from 3-5 man-days to 0.5 hours , requiring only 3 lines of core code . Both capabilities are implemented in ArkTS native with zero C++ dependencies .

These open source libraries enable more developers to conveniently reuse capabilities and collaborate, boosting HarmonyOS development efficiency.

Conclusion

From full-chain AI-assisted R&D tool co-construction, to hardcore technical optimizations in startup, networking, and GPU, to the landing of open source libraries across live streaming, image loading, and near-field sharing scenarios — every partner code contribution and every open source capability refinement jointly draws a blueprint for a more efficient, smoother, and smarter digital world. Each developer's shared technical capability takes root and sprouts in the HarmonyOS ecosystem. Thanks to all partners for their joint innovation and open sharing on HarmonyOS soil; looking forward to more developers and enterprise partners joining the HarmonyOS ecosystem to co-create infinite possibilities.