Helios: Open-Source 14B Video Model Generates 60+ Second Videos at 19.5 FPS on a Single H100 — First Real-Time Long-Video Model

The Decoder·medium signal

Helios, built by ByteDance, Peking University, and Canva, is the first 14B video model to hit near-real-time inference (19.5 FPS) on a single H100 while producing videos longer than 60 seconds. It achieves this without KV-cache, sparse attention, or quantization — instead using hierarchical memory compression (8x token reduction) and 3-step distillation. Text-to-video, image-to-video, and video-to-video are unified in one architecture. All code and weights are Apache 2.0 on GitHub and Hugging Face.