Pang Tianyu, Tencent's head of multimodal reinforcement learning technology for the Hunyuan model family, just unveiled HyCreator and opened up early access requests. Think of HyCreator as an "agent harness" for long-form video: a framework that strings together models, tools, and production workflows to generate roughly 10-minute films with zero human intervention. If you want to jump in and tweak things mid-production, you can switch over to real-time interactive editing at any point.
The official site already has a handful of finished pieces, with the longest running 10 minutes and 27 seconds. There are also examples at 9:45 and 9:59. Better yet, each project includes the Agent's full generation trace. In one public case, Pro mode executed 13 steps, produced 45 intermediate outputs, and ran checks on clip quality, shot transitions, and overall coherence before stitching everything into a final film.
HyCreator ships with two finishing modes: Lite and Pro. What video models are running underneath, how long a full generation takes, and what it costs — none of that's been disclosed yet.