Open video models have tended to force a choice between quality and hardware you can actually own. HunyuanVideo-1.5 is Tencent’s attempt to narrow that gap: 8.3 billion parameters, weights and inference code on GitHub and Hugging Face, and a documented minimum of 14GB of GPU memory with offloading.
| Repository | Tencent-Hunyuan/HunyuanVideo-1.5 |
| Maintainer | Tencent Hunyuan |
| Primary language | Python |
| GitHub stars | 4,551 (as of 15 September 2026) |
| Last push | 10 April 2026 |
| License status | Unverified — see below |
What it does
HunyuanVideo-1.5 generates video from text prompts (T2V) or from a starting image (I2V). The repository’s release history since the November 2025 launch shows steady additions:
- Cache-based acceleration (DeepCache, TeaCache, TaylorCache), added in late November 2025.
- A 480p image-to-video step-distilled model (December 2025) that runs in 8 or 12 steps. The README reports end-to-end generation within 75 seconds on a single RTX 4090, a 75% reduction.
- Training code and a LoRA fine-tuning script, using the Muon optimizer the team open-sourced alongside it.
- FP8 GEMM inference (December 2025), plus official Hugging Face Diffusers support and a ComfyUI usage guide.
The README also lists community integrations, including LightX2V and Wan2GP, which it says can run the model with as little as 6GB of VRAM.
Who it’s for
- Creators and developers who want a self-hosted video model on a single high-end consumer GPU.
- Teams that plan to fine-tune a video model with LoRA on their own footage.
- ComfyUI and Diffusers users, who have official integration paths.
Getting started
The README’s system requirements: an NVIDIA GPU with CUDA, 14GB minimum GPU memory with model offloading enabled, Linux and Python 3.10+. Setup is a clone, pip install -r requirements.txt, and optional attention libraries (Flash Attention, Flex-Block-Attention for sparse attention, or SageAttention) for faster inference. Weights come from Hugging Face; the step-distilled model is enabled with --enable_step_distill when running generate.py.
License status
Unverified. GitHub could not match this project’s license file to a standard open-source license (it reports NOASSERTION). The Signal has not reviewed the terms. Read the license file, and the terms on the Hugging Face model page, before any commercial use, redistribution or fine-tuning. Custom model licenses often include conditions a standard open-source license doesn’t.
Limitations and open questions
- The open-source plan in the README still lists “Release all model weights (Sparse attention, distill model, and SR models)” as not done.
- The official setup is Linux-only.
- The 14GB figure assumes offloading, which the README notes is slower than running fully on the GPU.
- Speed figures are Tencent’s own; we have not benchmarked them.
Sources
Related: ComfyUI-LTXVideo · Video Generation.
This profile was drafted with AI assistance from the project’s public repository and reviewed by The Signal’s editors before publication. See our AI Disclosure.
