HunyuanVideo vs Wan2.2: a reference framework against a permissive model suite
Both are open text-to-video research releases with weights you run yourself, and they make different bets. HunyuanVideo is Tencent's original framework, one model with pinned dependencies, a licence the repository itself declares as unclassified, and the largest surrounding ecosystem of community work. Wan2.2 is the Wan-AI suite, a Mixture-of-Experts architecture with a 5B hybrid that runs 720P at 24fps on a consumer card, plus speech-to-video and character animation models, under Apache-2.0. Choose HunyuanVideo for the reference implementation and the forks around it, and Wan2.2 for a permissive licence and a hardware path that fits a single GPU.
At a glance
| Project | Tencent-Hunyuan/HunyuanVideo | Wan-Video/Wan2.2 |
|---|---|---|
| Licence | Custom licenceCustom licence: read the LICENSE file | Apache-2.0Permissive: commercial use allowed |
| Maintenance | Commits in the last six monthsLast push October 8, 2026 | Commits in the last six monthsLast push September 21, 2026 |
| Language | Python | Python |
| GitHub stars | 12,738 | 18,449 |
| Read more | Our analysisGitHub | Our analysisGitHub |
Which one to choose
Choose HunyuanVideo if you want the original reference implementation, the exact model definition its paper describes, and the widest set of community alternatives to it. The repository ships PyTorch model definitions, pre-trained weights and inference and sampling code, integrates with Diffusers and ComfyUI, and has FP8 weights, a GGUF version, cache-based acceleration, a distilled model, sliding tile attention, a video length extrapolation method and a low-memory variant built on top of it. Choose it when you have the GPU memory for the full checkpoint, when you can pin torch==2.6.0 and match your CUDA build to it, and when you have read LICENSE.txt and the Notice for the weights.
Choose Wan2.2 if the licence is the first question, or if your hardware is one consumer GPU rather than a rack. It is Apache-2.0, it introduces a Mixture-of-Experts architecture that raises model capacity at the same computational cost, and its TI2V-5B hybrid generates 720P at 24fps from text or image prompts using a VAE with a 16x16x4 compression ratio, on cards such as a 4090. It also ships S2V-14B for audio-driven video and Animate-14B for character animation and replacement. Choose it when you need those specialist models or when Apache-2.0 is what your legal team needs to see.
One framework against a suite
HunyuanVideo is a framework in the original research sense. Its README says the repository contains PyTorch model definitions, pre-trained weights and inference and sampling code for the paper exploring HunyuanVideo, and the project page carries more visualisations. The architecture is one diffusion video model, and the news list records the sequence plainly: inference code and weights on Dec 3, 2024, parallel inference code powered by xDiT on Dec 7, integration into Diffusers on Dec 17, FP8 model weights to save GPU memory on Dec 18, a benchmark release in January 2025, an image-to-video model in March, a multimodal-driven customised generation architecture in May, an audio-driven human animation model in May, and in November 2025 a separate HunyuanVideo-1.5 repository described as a new foundation model. The direction of travel on this repository has been sideways into specialist models and a successor, not a stream of revisions to one checkpoint.
Wan2.2 is a suite, and its README organises itself as a list of innovations. The first is an Effective MoE Architecture, a Mixture-of-Experts design that separates the denoising process across timesteps with specialised expert models, which the project says enlarges overall model capacity while maintaining the same computational cost. The second is curated aesthetic data with detailed labels for lighting, composition, contrast and color tone. The third is a larger training set than Wan2.1, which the README puts at more images and more videos, and it claims top performance among open and closed models, which is the project's own claim rather than a measured result. The fourth is an Efficient High-Definition Hybrid TI2V: a 5B model built with the Wan2.2-VAE at a 16x16x4 compression ratio, supporting text-to-video and image-to-video at 720P and 24fps, and able to run on consumer graphics cards such as a 4090.
The second project's news list shows the same pattern at a different stage: weights and inference code in July 2025, then Diffusers and ComfyUI integration, then S2V-14B for audio-driven cinematic video with CosyVoice for speech synthesis, then Animate-14B for character animation and replacement, and that model into Diffusers by November 2025. That is a suite being filled in.
Weights, licences and what you can ship
The licence difference here is not a detail, and it is not symmetric.
HunyuanVideo's repository is listed under a licence that GitHub cannot classify, with the LICENSE file named as the place where the terms are. Our analysis is blunter: the repository declares NOASSERTION and the text-to-video weights carry their own terms, which is why it tells you to read LICENSE.txt and Notice before committing. A team shipping generated video in a product has to resolve that question with a lawyer rather than with a badge, and the answer may differ between the code and the weights.
Wan2.2 is Apache-2.0, which permits commercial use, modification and distribution with attribution and patent terms, and our analysis names the permissive licence as one of the reasons the specialist speech-to-video and character animation capabilities are worth having. Nothing in that sentence is a technical property of the model, and it is still often the property that decides.
The download paths differ too, and both are documented in ways worth reading before you plan capacity. HunyuanVideo's analysis sends you to ckpts/README.md for the download instructions, which is where a large checkpoint is fetched from. Wan2.2's README links its model weights on Hugging Face and its presence on ModelScope, and its specialist models are separate downloads for S2V-14B and Animate-14B.
There is a second-order point that follows from the ecosystem rather than the licence. HunyuanVideo has a GGUF version and an FP8 weight release, which change what has to be downloaded and how it runs. Quantised and lower-precision variants of an open checkpoint are what make a model reachable on a machine that cannot hold the full thing, and on that axis the first project has more roads than the second.
Hardware: the full checkpoint against a 5B hybrid
This is where the two are least interchangeable.
HunyuanVideo is built for hardware you have to think about. Our analysis describes it as built for people with serious GPU memory who want the model on their own hardware rather than a hosted playground, and it makes adoption conditional on having the GPU memory for the full checkpoint. The project anticipated this: the Dec 18, 2024 release of FP8 model weights is explicitly described as a way to save GPU memory, and the community list includes a GGUF version, a consistency distilled model, sliding tile attention, cache-based acceleration, sparse video generation, and a project described as a GPU poor version. Every one of those exists because the unmodified checkpoint does not fit or does not run fast enough for a lot of people.
Wan2.2 was designed with a smaller door in mind. The 5B hybrid is described as one of the fastest 720P at 24fps models available, and the README states it can also run on consumer-grade graphics cards like a 4090. Our analysis says the same thing from the other direction: consumer GPU users will find the TI2V-5B the most practical entry point. The MoE design is relevant here too, since our analysis reads it as the 14B models carrying more capacity without proportionally more compute, which is a way of keeping a larger model affordable rather than a way of making a small one capable.
The dependency picture follows the hardware. For HunyuanVideo, our analysis tells you to confirm your CUDA build matches torch==2.6.0 and that you are comfortable with pinned PyTorch dependencies, and it rules the project out for anyone who needs a one-command CPU install. For Wan2.2, our analysis tells you to check that the target GPU can handle the model's VRAM requirement and to confirm that flash_attn builds on your CUDA version, which is the same class of problem with a different library named. Neither project gives you a CPU path, and both need a build step that can fail on an unusual CUDA version.
The surrounding ecosystem and how you would run it
If you are not going to write the sampling code yourself, the ecosystem does the work. HunyuanVideo's community list is the longer one and it is worth reading as a menu of things you would otherwise build: FP8 inference with video-to-video and image-to-video generation through a ComfyUI wrapper, native ComfyUI support, a consistency distilled model with sliding tile attention, a GGUF and quantisation version, a free enhancement pass, cache-based acceleration, a low-memory variant, video length extrapolation from Tsinghua University, a keyframe control LoRA, and a packing of input frame context in next-frame prediction models. Diffusers carries a HunyuanVideo pipeline, and a prompt rewrite model is published alongside it. There is also a parallel inference path through xDiT, released in December 2024.
Wan2.2's integrations are fewer but deeper into the standard tooling: T2V-A14B, I2V-A14B and TI2V-5B are in Diffusers, Wan2.2 is in ComfyUI, there is a Hugging Face space running the TI2V-5B model, and the specialist models have their own spaces and are on ModelScope as well as Hugging Face. The specialist models are the part you cannot assemble from community work: S2V-14B is an audio-driven cinematic video generation model with its own technical report, paired with CosyVoice for text-to-speech synthesis in the speech-to-video task, and Animate-14B is a unified model for character animation and replacement that replicates movement and expression as a whole.
So the operational question is short. If you want to run someone else's sampling code over a checkpoint, the first project gives you more variants to choose from. If you want a capability that only exists as a trained model, the second project gives you two of them, under a licence your legal team can read in a minute.
Where HunyuanVideo falls short
The licence is the first limit, and it is a real one. Our analysis tells you to read LICENSE.txt and Notice because the repository declares NOASSERTION and the weights carry their own terms, and a project whose own repository cannot state a licence in a standard form is asking you for legal work before you start.
Second, the pinned stack. torch==2.6.0 with a matching CUDA build is a constraint, not a default, and our analysis makes comfort with pinned PyTorch dependencies part of the adoption condition. Anyone whose environment is on a different torch has to decide whether to build around it or to move.
Third, the hardware. There is no CPU path and no one-command install, which our analysis states plainly by telling you not to adopt the project if you need either. A team without a large GPU, or without the patience to run a quantised or distilled variant someone else maintains, is not served here.
Fourth, the direction of the project. The README's news list ends with a separate HunyuanVideo-1.5 repository in November 2025, described as a highly efficient and powerful new foundation model, and the last recorded push to this repository is 2026-06-29. The repository is not archived and that date is inside the last six months, so this is not a dead project. It is a project whose new capability is arriving next door, which means a team that wants the current state of the art from this line of work is looking at the 1.5 repository and treating this one as the reference and the ecosystem base.
Where Wan2.2 falls short
The capacity story cuts both ways. Our analysis says the MoE architecture lets the 14B models carry more capacity without proportionally more compute, and it also says consumer GPU users will find the TI2V-5B the most practical entry point. Those two sentences describe two different products, and the second one is the smaller. A team that wants the 14B models is back to reading VRAM numbers, and the design that makes them affordable is the same design that makes them awkward to reason about when you are debugging a specific sample.
Second, the build. Our analysis tells you to confirm that flash_attn builds on your CUDA version, which is a dependency with a known habit of not building, and to check the VRAM requirement against the target GPU. This is a normal video project requirement, and it is still a step that fails on a machine nobody tested.
Third, one requirements file per task. Our analysis tells you to review which requirements file, base, s2v or animate, matches the tasks you need, which means the suite is installed as several environments rather than one. That is a reasonable design for models that share little, and it is more container work than a single pip install.
Fourth, and this is a judgement about the ecosystem rather than a defect: the community work listed for this project is smaller, and much of the practical energy around video generation is pointed at quantisation, caching and distillation work that this project does not carry. The models are Apache-2.0, which is the right starting position for that work, but the work itself is someone else's to do.
Maintenance and the choice for concrete situations
HunyuanVideo's last recorded push is 2026-06-29 and Wan2.2's is 2026-09-21. Neither repository is archived and neither has GitHub releases, so in both cases you are tracking a branch rather than pinning a version, and the weights are versioned separately on Hugging Face. Wan2.2 is the more recently touched of the two, and its README's news list is the more active, with integration work recorded as recently as November 2025.
The licences, meanwhile, are the sharpest difference. Apache-2.0 is a licence you can hand to a legal team and get an answer from. An unclassified licence with separate weight terms cannot be handled that way, and our analysis makes reading both files part of the commitment rather than an afterthought.
The situations below are clear.
You need 720P at 24fps on one consumer GPU from a text or image prompt. Wan2.2, because the TI2V-5B hybrid with the Wan2.2-VAE is built for that card and that resolution.
You need speech-to-video with synthesised narration. Wan2.2, through S2V-14B and the CosyVoice speech synthesis support recorded in September 2025.
You need character animation or replacement with movement and expression replication. Wan2.2, through Animate-14B, which is also in Diffusers.
Commercial use where the licence has to be readable in a minute. Wan2.2.
You want the model that has the most community variants around it: quantised, distilled, cached, low-memory, keyframe-controlled, length-extrapolated. HunyuanVideo, and the community list is the argument.
You want a parallel inference path, or FP8 weights for memory. HunyuanVideo, both released in December 2024.
You want the exact implementation behind a specific paper, to build on. HunyuanVideo, which is what our analysis calls the reference implementation rather than a wrapper.
You want the newest foundation model in this line of work. HunyuanVideo-1.5, which the README announces as a separate repository rather than as part of this one.
Before committing, run the checks each project's analysis names. For HunyuanVideo, read the download instructions in ckpts/README.md, confirm your CUDA build matches torch==2.6.0, and read LICENSE.txt and Notice. For Wan2.2, check the VRAM requirement against your GPU, confirm flash_attn builds on your CUDA version, and pick the requirements file that matches your task, base, s2v or animate.
Bottom line
HunyuanVideo is the right choice when you want the reference implementation and the community variants built on top of it, because the repository ships the model definitions, the weights and the sampling code and the ecosystem around it covers memory, speed and format problems in ways you would otherwise solve yourself; Wan2.2 is the right choice when the licence must be Apache-2.0 or the hardware is a single consumer GPU, because the TI2V-5B hybrid runs 720P at 24fps on a card like a 4090 and the suite ships speech-to-video and character animation models nobody else publishes under a permissive licence. Verify first, for HunyuanVideo, that torch==2.6.0 matches your CUDA build and that you have read LICENSE.txt and the Notice for the weights, and for Wan2.2 that your GPU meets the VRAM requirement, that flash_attn builds, and that you have installed the requirements file matching your task. The judgement to take away is that this is not a quality contest between two models, since a permissive licence and a card you already own are operational facts while output quality is something you have to look at, so decide by those two facts first and judge the samples afterwards.
Sources
Was this useful?