ONLINEAGENT_OPS 2026.Q3 HOME HUBSMODELSBUILDINGCOMPANIESFILMMAKINGSECURITYLIVING WITH ITSKILLSALL HUBS, ONE TERMINAL → ARTICLES CRAFT RECORD BLOG TOOLS SKILLS SEARCH /
HOME›ARTICLES›CAN YOU RUN CHINESE AI VIDEO MODELS ON YOUR OWN
ARTICLES · GUIDEINTERMEDIATE

Can You Run Chinese AI Video Models on Your Own Computer? Wan, HunyuanVideo and MiniMax H3

Yes for Wan 2.2, HunyuanVideo 1.5 and MiniMax H3; no for Seedance, Kling and Wan 3.0. Graphics memory needed and licence limits by region, from the makers.

READ11 min
WORDS2,201
SECTIONS7
TYPEGUIDE
CHECKED27 OCT 26

Yes, for some of them. Alibaba’s Wan 2.2, Tencent’s HunyuanVideo 1.5 and MiniMax’s H3 publish files you can download and run yourself. Seedance, Kling and Alibaba’s newest Wan 3.0 do not. Two of the three licences also exclude readers in some regions, so read the licence before you download.

The smaller Wan 2.2 model needs a graphics card with 24 GB of video memory by the Wan project’s own instructions, HunyuanVideo 1.5 asks for at least 14 GB with a memory-saving setting turned on, and MiniMax gives no minimum for H3.

TL;DR — THE SHORT VERSION
  • Downloadable: Wan 2.2 (Alibaba), HunyuanVideo 1.5 (Tencent), MiniMax H3. Not downloadable: Seedance 2.5, Kling, Wan 3.0.
  • Memory (VRAM): Wan 2.2’s 5-billion-parameter model, 24 GB in the README and 8 GB with offloading in ComfyUI’s guide; HunyuanVideo 1.5, 14 GB minimum with offloading; H3, no minimum stated.
  • Licences: Wan 2.2 is Apache 2.0 with no region limit. HunyuanVideo 1.5 excludes the EU, UK and South Korea. H3 excludes the EU, UK, South Korea and the United States unless MiniMax grants you a licence.
  • Easiest first try: Wan 2.2’s small model, this site’s reading, because its licence is the plainest and its memory need is the lowest.
◈ IN PLAIN TERMS

Open weights means the company publishes the model’s files, so you can run it on your own computer instead of through its app. VRAM is the memory on your graphics card; the model has to fit in it to run at a sensible speed. Offloading means parking part of the model in the computer’s ordinary memory, which saves VRAM and costs speed. A checkpoint is one downloadable version of a model.

Which Chinese AI video models can you download?

The plain point: three families publish files today. Most of the newest, best-known names stay closed.

Downloadable Chinese video models, from each maker’s own pages (read 27 Oct 2026)
ModelWhat you can runLicence and regionMemory figure from the maker
Wan 2.2 (Alibaba)Makes: text-to-video and image-to-video at 480p or 720p; a 5B model at 720p, 24 frames a second; speech and character-animation models.Licence: Apache 2.0. No region limit.5B model: at least 24 GB. 14B models: at least 80 GB on one card, as the README’s commands say.
HunyuanVideo 1.5 (Tencent)Makes: text-to-video and image-to-video, 480p and 720p, from an 8.3-billion-parameter model.Licence: Tencent Hunyuan Community. Not for use in the EU, UK or South Korea.Minimum: 14 GB with offloading on.
MiniMax H3Makes: video with sound, about 768p, up to about 15 seconds.Licence: MiniMax H3 Community. Not for use in the EU, UK, South Korea or US without MiniMax’s licence.Minimum: not stated; builds are 16 GB to 66 GB of files.
Wan 3.0, Seedance 2.5, KlingNo download. Used through apps or paid APIs.Closed.n/a

Where each line comes from. Wan 2.2: the project’s README says the 5B model “supports both text-to-video and image-to-video generation at 720P resolution with 24fps and can also run on consumer-grade graphics cards like 4090,” and for its command, “This command can run on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU).” The same README says the 14B text-to-video command “can run on a GPU with at least 80GB VRAM,” with extra options if you run out of memory.1 HunyuanVideo 1.5: Tencent says it “delivers top-tier quality with only 8.3B parameters” (the company’s claim) and lists “Minimum GPU Memory: 14 GB (with model offloading enabled).”2 H3: ComfyUI’s guide says H3’s open weights “target about 1 megapixel (768px short edge) at 24fps for up to ~15 seconds,” and that “2K is a hosted-service output.” MiniMax’s model card gives no minimum memory; its own server example for H3 sets four GPUs.3 The closed models are covered on Chinese AI video models compared, which records Seedance 2.5 and Kling as closed (read 1 Oct 2026).

Wan 3.0 is not downloadable. On 27 Oct 2026 the public listing for Alibaba’s Wan-AI organisation on Hugging Face (the site where makers publish model files) held 27 models; the newest, dated 6 August 2026, are Wan 2.2 Animate versions, and none is named for Wan 2.5, 2.7 or 3.0. Alibaba says Wan is “Alibaba Cloud’s generative video model family” and that Wan 3.0 “is available through Alibaba Cloud Model Studio and supported international channels.” Treat a site offering a “Wan 3.0 download” with suspicion unless Alibaba’s own pages link to it.4

TAKEAWAY

Downloadable does not mean newest. The models you can run at home are Wan 2.2, HunyuanVideo 1.5 and H3; the newest Wan stays online only.

How much graphics memory do you really need?

The plain point: the maker’s figure is the floor for one setup, not a promise. Tools that move parts of the model out of VRAM can lower it, at the cost of speed.

Wan 2.2’s own README gives 24 GB for the small model, yet ComfyUI’s Wan 2.2 guide says “The Wan2.2 5B version should fit well on 8GB vram with the ComfyUI native offloading.” Both can be true: they describe different tools and settings.5 ComfyUI’s H3 guide gives file sizes for five H3 builds, from 16 GB for the smallest to 66 GB for the full unquantized one, and says “Pick the smallest build that fits in VRAM.” A build’s file size is not the same as the memory it uses while running, so treat those numbers as a guide to download size.3

New to the word? What is VRAM? explains it with a worktop comparison, and what is quantization? covers the smaller builds.

What it feels like to run one. Expect long first downloads (ComfyUI says tens of GB for H3) and some set-up work.

What do the licences let you do?

The plain point: Wan 2.2 is the easy one. The other two limit where you may use them, and neither lets you use their output to improve another AI.

Wan 2.2. The README says “The models in this repository are licensed under the Apache 2.0 License. We claim no rights over the your generated contents,” and ComfyUI’s guide adds that the Wan 2.2 models “are based on the Apache 2.0 open source license and support commercial use.” There is no region carve-out in the licence statement.6

HunyuanVideo 1.5. Tencent’s licence opens in capitals: “THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA” and is limited to the rest of the world. It also says you must not use the models “or any Output or results” there, must not use them “to improve any other AI model,” and that a company with more than 100 million monthly active users must ask Tencent for a licence. On output it says “Tencent claims no rights in Outputs You generate.”7

MiniMax H3. The licence covers “worldwide, excluding the Excluded Territories,” which it defines as “the European Union, the United Kingdom, the Republic of Korea and the United States of America.” It forbids use of the model “or any of their Outputs or results outside the Applicable Territory,” forbids using them to improve another AI model, requires a commercial product to “prominently display ‘MiniMax H3’”, and asks businesses above a yearly revenue limit it states to get MiniMax’s written authorisation first.8 The model card points people in the United States, EU, UK and South Korea to an application form, and its licence Q&A says organisations there “can apply for a formal license,” while the H3 API is “Globally available with built-in safeguards and responsible-use controls.”9

Two things in the H3 papers do not agree. First, the Q&A page’s heading says the licence is “currently limited to the EU, UK, South Korea, and US,” the opposite of the licence text, which excludes them. This page follows the licence text. Second, ComfyUI’s H3 guide says “Commercial use of locally generated outputs requires a MiniMax commercial license, available through Comfy, the only official reseller,” while the licence text itself only asks for separate authorisation above its revenue limit. If you plan to sell what you make, ask MiniMax which applies to you.10 For the full open-weights picture, including the other licence types, see AI labs outside the US and their open-weights licences.

TAKEAWAY

Look at where you live before you download. Wan 2.2 has no region limit; HunyuanVideo 1.5 and H3 do.

How do you actually run one?

The plain point: the usual route is ComfyUI, a free program that runs models through a diagram of connected boxes. You pick a ready-made workflow, let it fetch the files, and press run.

The Wan 2.2 README says it “has been integrated into ComfyUI” and Diffusers, a library for running models in code.11 Tencent says ComfyUI supports HunyuanVideo 1.5.12 For H3, ComfyUI’s steps are: update ComfyUI, then “Template Library > Video > pick a MiniMax H3 workflow,” then “Download models from the pop-up.”3 This site’s ComfyUI explained says who needs it, and warns that custom nodes are code from strangers.

  1. Check the licence for your country. It takes two minutes and decides whether you may start.
  2. Find your graphics card’s VRAM. Pick the smallest model that fits.
  3. Download only from the maker’s page. The README links above lead to Hugging Face pages under the makers’ own names.
  4. Start with a short, low-resolution clip. Longer, sharper clips need much more memory and time.

Steps 1 to 4 are this site’s own ordering, not a vendor procedure.

Is running it yourself more private?

The plain point: the model’s files run on your machine, so the maker’s app is not in the loop. That is this site’s reasoning, not a promise from any company.

A company’s privacy policy governs its own app and website, not weights you run on your own hardware; AI labs outside the US makes the same point and records that lookalike repositories using famous model names exist. Download from the maker’s own page. What the apps do with your uploads is covered in are Chinese AI video generators safe? A local run also sidesteps no licence term: the licences above still bind you.

Questions people ask

Can Wan 2.2 run on an 8 GB graphics card?

ComfyUI says the 5B version “should fit well on 8GB vram” with its native offloading. The Wan README asks for 24 GB for its command. The 14B models need far more on one card (source above).

Is Wan 3.0 open source?

No official download was found on 27 October 2026. Alibaba offers it through Alibaba Cloud Model Studio and “supported international channels” (source above).

Can I use MiniMax H3 in the United States?

The open weights licence excludes the United States unless MiniMax grants you a licence through its application form. The hosted H3 API is a separate route that MiniMax says is available globally (sources above). Read the licence yourself before relying on this summary.

Do the downloads make sound as well as pictures?

H3 does: ComfyUI says it generates “video with native stereo audio.” This page did not check the others.3

Where to go next

QUICK CHECK · 3 QUESTIONS

Pick an answer to see if you got it. Nothing is recorded.

1. Which Chinese video models does the page say you can download and run yourself?

  • Seedance 2.5, Kling and Wan 3.0
  • Wan 2.2, HunyuanVideo 1.5 and MiniMax H3
  • Only MiniMax H3
Show the answer

Answer: Wan 2.2, HunyuanVideo 1.5 and MiniMax H3. Wan 2.2, HunyuanVideo 1.5 and MiniMax H3 publish downloadable files. Seedance, Kling and Wan 3.0 are used through apps or paid APIs.

2. What do the HunyuanVideo 1.5 and MiniMax H3 licences have in common?

  • Both are Apache 2.0 licences
  • Both limit use to certain parts of the world
  • Both ban commercial use
Show the answer

Answer: Both limit use to certain parts of the world. HunyuanVideo 1.5 excludes the EU, UK and South Korea. H3 excludes the EU, UK, South Korea and the United States unless MiniMax grants a licence.

3. What does offloading do?

  • It deletes part of the model to save space
  • It makes the video longer
  • It parks part of the model in ordinary memory, which saves graphics memory but costs speed
Show the answer

Answer: It parks part of the model in ordinary memory, which saves graphics memory but costs speed. The page explains that offloading moves part of the model out of VRAM, so a smaller card can run it more slowly.

SOURCES
  1. 1Wan-Video, Wan2.2 README on GitHub, read at source 27 Oct 2026“This model supports both text-to-video and image-to-video generation at 720P resolution with 24fps and can also run on consumer-grade graphics cards like 4090.”; for TI2V-5B, “This command can run on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU).”; for T2V-A14B, “This command can run on a GPU with at least 80GB VRAM.” and “If you encounter OOM (Out-of-Memory) issues, you can use the `--offload_model True`, `--convert_model_dtype` and `--t5_cpu` options to reduce GPU memory usage.” ↑
  2. 2Tencent Hunyuan, HunyuanVideo-1.5 README on GitHub, read at source 27 Oct 2026“HunyuanVideo-1.5 is a video generation model that delivers top-tier quality with only 8.3B parameters, significantly lowering the barrier to usage.”; “Minimum GPU Memory: 14 GB (with model offloading enabled)”; “The memory requirements above are measured with model offloading enabled.” ↑
  3. 3ComfyUI, MiniMax H3 guide, read at source 27 Oct 2026“Open weights target about 1 megapixel (768px short edge) at 24fps for up to ~15 seconds. 2K is a hosted-service output and needs a separate upscale pass in ComfyUI.” MiniMax, MiniMax-H3 model card, read at source 27 Oct 2026: the “Sglang Deployment” example uses “--num-gpus 4”; no minimum-memory line found on the card. ↑a“Pick the smallest build that fits in VRAM.”; the table lists pruned_w6a8 at “16 GB”, pruned_int8_convrot and pruned_fp8_scaled at “21 GB”, pruned_bf16 at “40 GB” and bf16 at “66 GB”; “Download models from the pop-up (tens of GB; allow time and disk space)”. ↑b“Template Library > Video > pick a MiniMax H3 workflow”; “Download models from the pop-up (tens of GB; allow time and disk space)”. ↑c“generates video with native stereo audio (voice, SFX, and music in a single pass)”. ↑d
  4. 4Hugging Face, public model listing for the Wan-AI organisation, read at source 27 Oct 202627 models, newest “Wan-AI/Wan2.2-Animate-2-14B-Distilled-Diffusers” created 2026-08-06. Alibaba Cloud, Wan3.0 at General Availability, read at source 27 Oct 2026: “Wan is Alibaba Cloud’s generative video model family. Wan3.0 is available through Alibaba Cloud Model Studio and supported international channels, with API pricing billed per generated second.” ↑
  5. 5ComfyUI, Wan2.2 video generation ComfyUI native workflow example, read at source 27 Oct 2026“The Wan2.2 5B version should fit well on 8GB vram with the ComfyUI native offloading.” ↑
  6. 6Wan-Video, Wan2.2 README, “License Agreement”, read at source 27 Oct 2026“The models in this repository are licensed under the Apache 2.0 License. We claim no rights over the your generated contents, granting you the freedom to use them while ensuring that your usage complies with the provisions of this license.” ComfyUI, Wan2.2 guide, read at source 27 Oct 2026: “The Wan2.2 series models are based on the Apache 2.0 open source license and support commercial use.” ↑
  7. 7Tencent Hunyuan, Tencent Hunyuan Community License Agreement (HunyuanVideo 1.5, release date 21 Nov 2025), read at source 27 Oct 2026“THIS LICENSE AGREEMENT DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA AND IS EXPRESSLY LIMITED TO THE TERRITORY, AS DEFINED BELOW.”; “You must not use, reproduce, modify, distribute, or display the Tencent Hunyuan Works, Output or results of the Tencent Hunyuan Works outside the Territory.”; “You must not use the Tencent Hunyuan Works or any Output or results of the Tencent Hunyuan Works to improve any other AI model”; “greater than 100 million monthly active users”; “Tencent claims no rights in Outputs You generate.” ↑
  8. 8MiniMax, MiniMax H3 Community License Agreement, read at source 27 Oct 2026“‘Applicable Territory’ means worldwide, excluding the Excluded Territories.” and “‘Excluded Territories’ means the European Union, the United Kingdom, the Republic of Korea and the United States of America.”; “You may not use, reproduce, modify, distribute, or display the MiniMax H3 Works or any of their Outputs or results outside the Applicable Territory.”; “You may not use the MiniMax H3 Works or any of their Outputs or results to improve any other artificial intelligence model”; “You shall prominently display “MiniMax H3” on the user interface of commercial product or service that uses MiniMax H3 or MiniMax H3 Works.”; section IV.1 requires separate written authorization above a stated yearly revenue figure. ↑
  9. 9MiniMax, MiniMax-H3 model card, read at source 27 Oct 2026“Application form(only for USA/EU/UK/South Korea)”. MiniMax, Q&A about License, read at source 27 Oct 2026: “Organizations in these regions can apply for a formal license.” and “API: Globally available with built-in safeguards and responsible-use controls.” ↑
  10. 10MiniMax, Q&A about License, read at source 27 Oct 2026, first heading“Why is MiniMax-H3’s open-weight license currently limited to the EU, UK, South Korea, and US?” ComfyUI, MiniMax H3 guide, read at source 27 Oct 2026: “Commercial use of locally generated outputs requires a MiniMax commercial license, available through Comfy, the only official reseller.” ↑
  11. 11Wan-Video, Wan2.2 README, read at source 27 Oct 2026“Wan2.2 has been integrated into ComfyUI”; “Wan2.2’s T2V, I2V and TI2V have been integrated into Diffusers”. ↑
  12. 12Tencent Hunyuan, HunyuanVideo-1.5 README, read at source 27 Oct 2026“ComfyUI supports HunyuanVideo-1.5 with various engineering optimizations for fast inference.” ↑

Read at source 27 Oct 2026: Wan-Video, Wan2.2 README; Hugging Face, Wan-AI model listing; Alibaba Cloud, Wan3.0 at General Availability; Tencent Hunyuan, HunyuanVideo-1.5 README and licence; MiniMax, MiniMax-H3 model card, licence and licence Q&A; ComfyUI, Wan2.2 guide and MiniMax H3 guide.

Reused from this site’s Chinese AI video models compared (read 1 Oct 2026): Seedance 2.5 and Kling are closed.

Not checked: any model’s real speed or quality on a home computer, who runs the GitHub project behind the Wan 2.2 README (Alibaba Cloud calls Wan its own model family), and the H3 papers’ two disagreements noted above. Memory figures are the makers’ and ComfyUI’s, and change with each tool release. The order of steps, the reading that Wan 2.2 is the easiest first try, and the privacy reasoning are this site’s own, not claims from these sources. This page is not legal advice.

Free printable AI cheat sheets, and AI news every week.

Sixteen one-page AI guides and checklists in one free PDF, no sign-up. And if you want to keep up, the newsletter covers what changed in AI each week, in plain English, with the sources.

SUBSCRIBE FREE ↗ GET THE CHEAT SHEET PACK
aipromptgeneer://$ about this siteAI PROMPT GENEERAI explained in plain English. Every claim links to its source, with the date it was checked.NEWSLETTERWhat changed in AI each week, in plain English.SUBSCRIBE FREE ↗AI PROMPT GENEER · MACHINE ASSISTANCE DISCLOSEDEVERY ARTICLE CARRIES ITS OWN CHECKED DATE