Helios is an open-source, real-time long video generation model from PKU-YuanGroup, the team behind Open-Sora Plan. It is a 14B diffusion transformer that achieves minute-scale, high-quality video synthesis at 19.5 FPS on a single H100 GPU (about 10 FPS on a single Ascend NPU). The headline result is notable because Helios reaches this without relying on the conventional long-video anti-drifting strategies or standard video acceleration techniques that most real-time systems depend on.

Minute-Scale Video Without Anti-Drifting Tricks

Most long-video generators fight temporal drift with mechanisms like self-forcing, error banks, keyframe sampling, or inverted sampling. Helios generates minute-scale videos with high quality and strong coherence without any of these, simplifying the pipeline while maintaining consistency over long durations. This makes it a clean baseline for studying long-horizon video generation.

Real-Time Speed Without Standard Acceleration

Helios also forgoes the usual acceleration toolbox — no KV-cache, causal masking, sparse or linear attention, TinyVAE, progressive noise schedules, hidden-state caching, or quantization — yet still delivers 19.5 FPS end-to-end on a single H100, with some users reporting over 20 FPS on stronger hardware. Instead, the authors introduce optimizations that raise both training and inference throughput while cutting memory, enabling image-diffusion-scale batch sizes during training and fitting up to four 14B models within 80GB of GPU memory.

Flexible Deployment and Low VRAM

Despite its 14B scale, Helios is engineered for accessibility. Group Offloading brings VRAM usage down to roughly 6GB, and Context Parallelism enables multi-GPU inference with Ulysses, Ring, Unified, and Ulysses-Anything attention variants. Community tutorials demonstrate installation on consumer-grade PCs and even 4K video generation, broadening its reach beyond datacenter hardware.

Broad Ecosystem Support

Helios launched with day-0 integration across the generative video ecosystem: Diffusers, SGLang-Diffusion, vLLM-Omni, and Huawei Ascend-NPU, plus Cache-DiT acceleration support. It ships three model variants — Helios-Base (best quality with v-prediction and a custom HeliosScheduler), Helios-Mid, and Helios-Distilled — covering text-to-video, image-to-video, video-to-video, and interactive generation. A Gradio demo, technical report on arXiv, and HuggingFace/ModelScope collections round out the release.

Assessment

Helios is a strong open-source contribution to real-time long video generation, pairing a genuinely fast inference path with permissive Apache-2.0 licensing and unusually broad framework support. Its willingness to drop common anti-drifting and acceleration tricks makes it both a practical tool and a useful research reference. The main caveats are hardware-dependent performance — real-world FPS varies with CPU, memory, and driver versions — and the fact that H100-class throughput remains out of reach for most consumer setups, where VRAM optimizations trade speed for accessibility.