Not sure which to join? Get 1:1 help →
Build real AI applications on Radeon GPUs and the open ROCm software stack. Three tracks: multimodal, agentic AI and physical AI, with community support from Dify and others and an emphasis on local deployment.
Browse similarOnline hackathonsAI hackathonsAI Agent hackathonsHackathons in August 2026
No preview · GitHubFrom product photos, brand assets and a spoken brief, generates three-part vertical ads; products and logos are composited from source files and every scene passes 11 checks before export. 1st place, Multimodal track.
An offline CLI tool that turns any video into an audio-described soundtrack so blind and low-vision viewers can follow films and shows, running fully on a Radeon GPU. 2nd place, Multimodal track.
A child speaks a story idea into the mic and about two and a half minutes later gets an illustrated, narrated, subtitled picture-book video and a printable PDF — all five models run locally. 3rd place, Multimodal track.
No preview · GitHubA local research environment for individual investors that uses multi-agent planning, controlled retrieval and 80 deterministic financial functions to turn public company filings into an auditable research process. 1st place, Agent track.
No preview · GitHubA coding assistant that runs entirely on a local Radeon GPU, with a CLI and a Flutter desktop app, powered by a HIP inference engine written from scratch. 2nd place, Agent track.
No preview · GitHubA privately deployed QA agent for customer service conversations that combines deterministic rules, semantic judgment, retrieved evidence and human review to flag compliance issues and follow up on tickets. 3rd place, Agent track.
Ports the PhysX 5 GPU rigid-body solver from CUDA to HIP so ManiSkill robot simulation runs on Radeon, then trains visual policies and deploys them to a real robot arm. 1st place, Physical AI track.
Real-scene 3DGS reconstruction, data collection, VLA post-training and closed-loop evaluation, all on a single Radeon GPU — demoed by a robot unloading a dishwasher. 2nd place, Physical AI track.
No preview · GitHubRecord an ordinary video of human movement; pose estimation, motion retargeting and reinforcement learning turn it into a control policy the Unitree G1 can execute. 3rd place, Physical AI track.
The full original listing below is in Chinese, as published in the Hack123 newsletter.
基于 Radeon GPU 与 ROCm 开放软件栈构建真实 AI 应用。
看点：多模态、Agentic AI、Physical AI 三赛道；Dify 等社区支持；强调本地部署
奖金：一等奖5,000美元、二等奖3,500美元、三等奖1,500美元
截止日期：2026年8月6日 11:59 PM（线上）
Summarized from the Chinese “Beginner Hackathon Series” newsletter. Eligibility, dates and prizes are subject to the organizer’s official page.