Helios

Helios is a breakthrough video generation model developed by researchers from Peking University, ByteDance, and Canva, released in March 2026. It is the world's first 14-billion-parameter video model to achieve real-time generation speeds on a single GPU.

**The key achievement**: Most video generation models require multi-GPU clusters and take minutes to render short clips. Helios generates minute-length, high-quality video at 19.5 FPS on a single NVIDIA H100 GPU — without relying on conventional acceleration techniques (distillation, anti-drifting tricks, or heavy quantization). On a single Ascend NPU, it achieves approximately 10 FPS.

**Technical highlights**: - **14B parameters** — Large enough for high visual quality, optimized for speed through architectural innovation - **Minute-scale generation** — Sustained, temporally consistent long-form video synthesis (not just 4-8 second clips) - **Real-time throughput** — ~19.5 FPS on H100, making it practical for interactive or streaming applications - **No conventional tricks** — Achieves its speed through fundamental architectural choices, not post-hoc acceleration hacks - **Matches strong baselines** — Quality is competitive with leading video models despite the speed advantage

**License and availability**: Helios is released under Apache 2.0, making it fully open for research and commercial use. Code, weights, and the paper (arXiv:2603.04379) are available on GitHub at PKU-YuanGroup/Helios.

Helios represents a fundamental shift in what's possible with video generation on consumer-accessible hardware — making real-time AI video generation practical for studios, interactive media, and developer applications.