Correction, 2026-10-02: this page read as if LTX-2.5 and Wan 2.2 were the whole open-weights field, and its shortlist left out MiniMax H3, whose weights MiniMax published on Hugging Face on 2026-07-28 (about 3.57 million downloads in the past 30 days, checked 2026-10-02). H3 is now in the shortlist with its main catch: its community licence does not cover the United States, the United Kingdom, the European Union or South Korea, so a reader there cannot use it under that licence. The shortlist also named LTX-2.5's licence the LTX-2 Community License; it is the LTX-2.x Community License Agreement. What the page says about Wan is unchanged: there are no Wan 2.5 or Wan 3.0 weights.
Correction, 2026-10-01: we called hosted Wan 2.2 on 8frame "exactly the model you would have self-hosted" and suggested running the same prompt through it. On 8frame, Wan 2.2 runs only from a start image or an input video, while the self-hosted weights also do text-to-video. For a text prompt on 8frame, the Wan model is Wan 3.0 (24 credits for 5 seconds at 480p). A later pass the same day removed an unmeasured quality comparison with Kling v3 and Veo 3.1, updated the LTX-2.5 download figure and added sources for the outside facts.
No. Wan 3.0 is API-only, with no weights published on Hugging Face, GitHub, or ModelScope. Neither is Wan 2.5, despite how often it gets described as open. The last Wan you can actually download is Wan 2.2, released under Apache 2.0, and that is still one real answer to "which open-weights video model can I run on my own GPU in 2026," though not the only one: the others are in the shortlist below.
TL;DR
- Wan 3.0: no public weights. Hosted only, through Alibaba Cloud Model Studio, fal (
alibaba/wan-3.0/*andalibaba/wan-3.0-prime/*) and Replicate (alibaba/wan-3). It is on the 8frame canvas from 24 credits per 5-second clip - Wan 2.5: also no public weights. It was the point where the Wan line went closed
- Wan 2.2: Apache 2.0, weights on Hugging Face, 720p at 24fps, 5-second clips, 24GB VRAM class GPU
- Hosted route if you do not have the card: Wan 2.2 at 50 credits per 5s clip on 8frame (from a start image or video; it does not take text alone there), Wan 2.5 at 102
How to check this yourself
The claim is easy to verify, which is why it is strange that so many pages get it wrong. One request lists everything the Wan team has published:
curl "https://huggingface.co/api/models?author=Wan-AI&sort=createdAt&direction=-1"
We ran that on 2026-08-28 and again on 2026-10-01, with the same result. Newest repositories, in order: Wan2.2-Animate-2-14B-Distilled-Diffusers and Wan2.2-Animate-2-14B-Diffusers (2026-08-06), Wan2.2-Animate-2-14B (2026-07-14), Wan-Dancer-14B (2026-07-10), then the 2025 Wan 2.2 and Wan 2.1 releases. No 2.5. No 3.0.
The pattern is worth naming, because it will repeat with the next release: Alibaba keeps publishing weights for the previous generation while the current flagship stays behind the API. Wan 2.1 and 2.2 are open. Everything above them is a product, not a download.
What "open" was buying you, and what replaced it
Self-hosting an open-weights video model was never about the model being better. It was about three things that hosted tiers take away: no watermark, no per-clip meter, and no cap on how many times you can iterate at 3am on a shot nobody is paying for yet.
LTX-2.5 and Wan 2.2 both still give you all three, and LTX-2.5 is the newer of the two. They are not the whole field: MiniMax H3 also has published weights, but its licence does not cover the United States, the United Kingdom, the European Union or South Korea (details in the shortlist below). The Wan2.2-TI2V-5B model card states 720p output at 1280x704 or 704x1280, 24fps, 5-second clips, a 24GB minimum VRAM requirement (an RTX 4090 class card), and generation in under 9 minutes on a single consumer GPU. That is a real production floor for b-roll, texture plates, and drafting.
What it does not give you is a 2026 model. Wan 2.2 is a 2025 release, capped at 720p and 5-second clips. Run your own prompt through it and a current hosted model before you plan a project around it, and use it for the shots where iteration count matters more than the last stretch of polish.
Correction, and why we are printing it
We previously described Wan 2.5 as an open-weights model across a number of posts, including the ones that rank. That was wrong, and it was wrong in the direction that costs a reader real money: someone with a 4090 who read our recommendation went looking for weights that do not exist. Wan 2.5 is offered through APIs only, and its weights have never been published on the Wan-AI Hugging Face account. We have corrected the claim wherever it appeared and the correct open pick is Wan 2.2.
If you need open weights, here is the actual 2026 shortlist
| Need | Model | License | Reality check |
|---|---|---|---|
| Current-generation output with audio | LTX-2.5 (Lightricks) | LTX-2.x Community License | Free commercially under $10M revenue, about 1.6 million downloads in the past 30 days |
| Video with stereo audio, outside the US, UK, EU and South Korea | MiniMax H3 (MiniMax) | MiniMax H3 Community License | Not licensed in the US, UK, EU or South Korea; 768p by default, 4 to 15 s per the card; commercial products must display "MiniMax H3"; about 3.57 million downloads in the past 30 days |
| The most permissive license | Wan 2.2 (TI2V-5B, T2V-A14B, I2V-A14B) | Apache 2.0 | 720p, 24fps, 5s, 24GB VRAM |
| Character animation from a driving video | Wan 2.2 Animate | Apache 2.0 | Newest Wan-AI uploads, 2026-08-06 |
| Stills and frames | FLUX open variants | Model-specific, check the card | Large community LoRA ecosystem |
| Veo, Kling, Sora, Seedance, Gemini Omni, Wan 2.5 and 3.0 | Not available | Closed | API-only, no weights published |
When hosted is the cheaper answer
Self-hosting is free per clip and expensive everywhere else: the card, the inference stack, the driver drift, and the hours. The break-even is lower than most people assume. On 8frame, Wan 2.2 runs 50 credits for a 5-second 720p clip and the $19 Starter plan carries 1,000 credits, which is 20 clips of the model you would have self-hosted, plus access to the closed models on the same canvas when a shot needs to actually look expensive. One difference: on 8frame, Wan 2.2 runs from a start image or an input video only, while the self-hosted weights include text-to-video; for a text prompt there, Wan 3.0 is 24 credits for 5 seconds at 480p.
If your monthly volume is under a few hundred clips and you do not already own the GPU, the hardware never pays back. If you are running thousands and have the card, self-hosting wins and always has.
FAQ
Will Wan 3.0 weights be released later? We found no announced date as of 2026-10-01. The precedent from 2.5 is that the flagship stays closed while the previous generation stays open, so the realistic expectation is Wan 2.2 remains the newest downloadable Wan for a while.
Can I use Wan 2.2 output commercially? Apache 2.0 is a permissive license, but read the model card for the specific repository you download since the Wan-AI org has published under more than one arrangement. On 8frame's paid plans, commercial use of output is cleared on every model.
Is Wan 3.0 on 8frame? Yes, since late August 2026, in both the base and Prime variants: 24 to 95 credits for a 5-second clip depending on resolution, or 46 to 189 on Prime. See Wan 3.0 pricing explained.
Sources
- Wan 2.2 input requirement on 8frame (start image or input video) and Wan 2.2 and Wan 3.0 prices: 8frame video generation code and tool registry, read 2026-10-01
- Wan-AI repositories, newest first (no Wan 2.5 or 3.0 weights): Hugging Face API, author Wan-AI, checked 2026-10-01
- Wan 2.2 TI2V-5B specs (720p at 1280x704 or 704x1280, 24fps, 5 s, 24GB VRAM, under 9 minutes on one consumer GPU): model card, checked 2026-10-01
- LTX-2.5 licence (LTX-2.x Community License Agreement; commercial use free under $10M revenue, per the model page on 2026-08-28) and downloads in the past 30 days (checked 2026-10-01): Hugging Face, Lightricks/LTX-2.5
- MiniMax H3: model card (768p default, stereo audio, 4 to 15 s), licence (dated 2026-08-02; excluded territories, "MiniMax H3" display) and Hugging Face API (created 2026-07-28, about 3.57 million downloads in the past 30 days), checked 2026-10-02
- Hosted Wan 3.0: Replicate, alibaba/wan-3 and fal, alibaba/wan-3.0-prime, checked 2026-10-01
The fastest way to find out whether open weights actually matter for your work is to run the same start image and prompt through Wan 2.2 and a closed flagship back to back (on 8frame, Wan 2.2 needs a start image). That takes one canvas and no GPU. Compare the field in best AI video generator 2026, or start on 8frame from $19/month.
Try it without signing up: the 8frame canvas opens on a real board. The canvas is free and unlimited; generation is paid from $19/month.