On HeliosBench, MovieRoll-Distilled achieves the highest total scores among compared methods across single- and multi-event evaluations up to 60s, while MovieRoll-Base stays competitive at 30s after training on videos of up to 15s (arXiv). The paper also says its KV-cache compression reduces cache memory by up to 78% while improving generation quality. The project has released training and inference code and model weights, but the licence excludes some territories and bars using outputs to improve other AI models (GitHub, Hugging Face, GitHub).
A base-to-distilled pipeline
MovieRoll is described by its authors as a base-to-distilled framework for long-form streaming video generation. It connects three stages: streaming fine-tuning, long-rollout distillation and long-horizon inference. For practitioners, the framing suggests the stages are meant to be used together rather than as separate tricks added to a short-clip model.
MovieRoll-Base is initialized from HunyuanVideo-1.5-480P-T2V (arXiv). According to the paper, MovieRoll-Base uses 6 full-history blocks and 48 sink–recent blocks. That split is the architectural basis for the memory design described next.
How HieraKV saves memory
The memory claim comes from HieraKV, a hierarchical KV-cache compression method that exploits layer-wise attention patterns. It keeps the full history in shallow layers and only the sink and recent history in deeper layers. The wording is “up to”, so the 78% figure for HieraKV describes the best case the paper reports, not a guaranteed saving in every setting.
The block counts for MovieRoll-Base line up with that design: a small number of full-history blocks and a much larger number of sink–recent blocks. Since HieraKV keeps the full history only in shallow layers and just the sink and recent history in deeper layers, the paper reports that it cuts cache memory by up to 78% compared with caching everything.
Reading the benchmark claim
The second point in the benchmark result is the more interesting one for long-form work: the base model holds up at twice the length of its 15s training clips. The paper credits generation beyond the training horizon to KV recomputation, which bounds context for both models, and to guidance annealing (AnnealCFG), which stabilizes MovieRoll-Base. The comparisons are the paper’s own reporting, so independent reproduction would be the check that matters.
Release history and licence terms
| Date | Entry, as written by the project | Source page | Read on |
|---|---|---|---|
| 2026-11-16 | Day-0 support for SGLang Diffusion and vLLM-Omni. | https://raw.githubusercontent.com/Tencent-Hunyuan/MovieRoll/HEAD/README.md | 2026-11-18 |
| 2026-11-16 | We release the MovieRoll code! | https://raw.githubusercontent.com/Tencent-Hunyuan/MovieRoll/HEAD/README.md | 2026-11-18 |
The project’s own release notes carry two entries, both dated 2026-11-16. One announces day-0 support for SGLang Diffusion and vLLM-Omni, and the other says the MovieRoll code was released. The shared date suggests the code and the serving-engine integrations arrived together. Support for two inference engines in the same release points to a focus on deployment, not only on the paper’s research results.
The GitHub README lists training and inference code and model weights as released. The licence applies only in a Territory that is worldwide excluding the European Union, the United Kingdom and South Korea, and grants no rights in those territories. It also forbids using the Tencent Hunyuan works or any output or results of them to improve any other AI model. Teams in the excluded regions cannot use the weights under this licence, and the output restriction rules out using generations as training data for other models.
How we built this
Release historyDated release entries published by the project
Dated entries read by our code from the project's own release notes, news list or changelog, newest first, with the page each was read from; the entry text is the project's own wording, not a check of what it says. Each entry is shown whole. A link in an entry is shown as its text; the source page has it.
Sources
The sources this article was written from. Links open the original page.
- arXivMovieRoll: A Systematic Framework for Rolling Out Long-Form Videosarxiv.org
- arXivMovieRoll: A Systematic Framework for Rolling Out Long-Form Videosarxiv.org
- GitHubhttps://github.com/Tencent-Hunyuan/MovieRoll: page linked by the paper as its resourceraw.githubusercontent.com
- Hugging Facehttps://huggingface.co/yunyangge/MovieRoll: page linked by the paper as its resourcehuggingface.co
- GitHubLicence or terms of use linked by https://github.com/Tencent-Hunyuan/MovieRollraw.githubusercontent.com


