Helios: a 14B video model that claims 19.5 FPS on one H100 without the usual anti-drifting or caching tricks

Minute-long clips without self-forcing, error banks or keyframe sampling

Long video models drift: after some seconds, faces and scenes degrade. Helios claims minute-scale, high-quality output while skipping the strategies that are normally used against drift. The README names them: self-forcing, error-banks, keyframe sampling and inverted sampling. None of the four are used.

The same list-of-omissions framing applies to speed. The README says Helios reaches 19.5 FPS in end-to-end inference on a single H100 without KV-cache, causal masking, sparse or linear attention, TinyVAE, progressive noise schedules, hidden-state caching or quantization. A single Ascend NPU is quoted at about 10 FPS.

These are claims about what the model avoids, and the repository gives the arXiv technical report (2603.04379) as the place for the method itself. If you are evaluating the approach rather than the checkpoint, read the paper before the code. The README describes the results and the omissions, not the architecture that replaces them.

Four 14B models fit in 80 GB because of training optimizations

The third highlight is about memory. The authors introduce optimizations that raise both training and inference throughput while cutting memory use, and they say this allows image-diffusion-scale batch sizes during training. The concrete figure given is up to four 14B models fitting within 80 GB of GPU memory.

The repository layout backs a training story. Next to infer_helios.py and app.py there is train_helios.py, a scripts/ directory, a tools/ directory, an eval/ directory, and the helios/ package. One configuration file is named explicitly in the news: scripts/training/configs/correct.yaml, introduced alongside fixes to the image-to-video train-inference inconsistency and to Easy Anti-Drifting.

That 2026-03-20 note matters for anyone fine-tuning. It says the authors identified, from issue #38, several ways to improve performance, and that the commits and correct.yaml hold the details. Training from a checkout older than that date means training with the inconsistency still in place.

Three checkpoints: Base, Mid and Distilled

On 2026-03-04 the project released training and inference code with weights for Helios-Base, Helios-Mid and Helios-Distilled. The weights are collected on Hugging Face and ModelScope, and a Gradio demo runs as a Hugging Face Space named Helios-14B-RealTime. The repository also ships app.py for running the demo yourself.

The README does not tabulate how the three variants differ in quality, step count or speed. Check the model cards before picking one. CNAPS.AI, an online inference platform, published separate tutorials for helios-base and helios-distilled.

The repository has no GitHub releases. Versioning is by commit and by news entry, so if you want reproducibility, record the commit hash you used along with the checkpoint name.

Group offloading and context parallelism trade speed for reach

Two features from 2026-03-08 stretch the hardware range. Group Offloading brings VRAM usage down to roughly 6 GB, according to the README, which opens the model to cards far below an H100. The README quotes the 19.5 FPS figure for a single H100 and gives no speed for offloaded runs.

Context Parallelism goes the other direction, spreading inference across several GPUs. The supported attention variants are Ulysses Attention, Ring Attention, Unified Attention and Ulysses Anything Attention. A fix to the parallel inference logits landed on 2026-03-06, with an example in the README. Anyone who tried multi-GPU inference before that date should update.

An unofficial community video, listed on 2026-03-24, covers installing Helios on a consumer-grade PC and generating 4K video. It is not an official resource, and the README labels it unofficial. Treat it as a pointer, not as a supported configuration.

The 19.5 FPS number depends on the CPU, memory and CUDA driver too

A note dated 2026-03-12 qualifies the headline figure. Real-time generation performance depends not only on the GPU but also on the CPU, memory, CUDA driver version and so on. A user on better hardware with a single H100 reported up to 20.89 FPS in issue #3, which shows the number moving upward. By the same logic, a slower host can pull it below 19.5.

That qualification has a practical edge. Buying or renting an H100 does not by itself guarantee the quoted speed. A cloud instance with a weak CPU or an older CUDA driver may fall short, and the README gives no minimum host specification.

Speed on Ascend is quoted separately, at about 10 FPS on a single NPU. The repository carries requirements_npu.txt next to requirements.txt, so the NPU install is a separate dependency set, not a flag on the default one.

A pinned stack: torch 2.10.0, diffusers 0.38.0, transformers 5.3.0

requirements.txt pins most of its heavy packages exactly: torch==2.10.0, torchvision==0.25.0, torchaudio==2.10.0, triton==3.6.0, kernels==0.13.0, diffusers==0.38.0, transformers==5.3.0, accelerate==1.12.0, deepspeed==0.18.4 and peft==0.18.1. Earlier diffusers and transformers pins are left in the file as comments. numpy is held below 2.0.0 with numpy<2.0.0.

Several entries have no version at all: opencv-python, gradio, moviepy, mpi4py, omegaconf and others. Those will resolve to whatever is current when you install. mpi4py needs a working MPI library to build, so a clean machine can fail at that line before anything else. An install.sh sits at the top level for the scripted route, though the README does not document how it is invoked.

Plan for a dedicated virtual environment. Exact pins on torch 2.10.0 and transformers 5.3.0 leave little room to share an environment with other projects.

Day-0 integrations: Diffusers, SGLang-Diffusion, vLLM-Omni and Cache-DiT

Beyond the repository's own scripts, Helios launched on 2026-03-04 with Day-0 support in Diffusers, SGLang-Diffusion and vLLM-Omni, each through a pull request linked from the README, plus Ascend NPU support. On 2026-03-06 Cache-DiT added support, offering cache acceleration and parallelism for Helios. Hugging Face Spaces gained Ahead-of-Time Compilation support on 2026-03-20.

Each route has a different trade. Diffusers is the familiar entry point for people already using it. SGLang-Diffusion describes itself as an inference framework with an end-to-end unified pipeline and an efficient scheduler loop. vLLM-Omni is described as a serving system that breaks complex architectures into a stage-based graph.

The repository is licensed Apache-2.0, per LICENSE.txt, and its last push was on 2026-08-24. A FAQ, tips and tutorials summary is pinned in issue #47, which is the place to look before opening a new issue.

Editorial conclusion

Use Helios if you have an H100 or an Ascend NPU, or a card with enough memory for group offloading, and want to study or fine-tune a minute-scale video model under Apache 2.0. Skip it if you need a pinned, released package, since the repository has no GitHub releases and requirements.txt pins torch 2.10.0 and transformers 5.3.0. Before planning anything around the 19.5 FPS figure, check your CPU, memory and CUDA driver version, because the authors say those matter as much as the GPU.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. PKU-YuanGroup/Helios on GitHub
4. Project website
5. README