“Wan 2.5 open source” is one of the most searched video-AI questions, and the web is full of wrong answers. Some sites claim you can download Wan 2.5, 2.6 or 2.7 from Hugging Face. You cannot.
This guide gives the straight answer for every Wan version up to Wan 3.0, shows where you can use the closed ones (and which are free), and walks through the best thing you can run at home today: Wan 2.2 in ComfyUI, with exact files, folders and VRAM needs.
- Is Wan 2.5 open source? No. Wan 2.5-Preview (24 Sep 2025) was only ever hosted. No weights, code or licence were released.
- Wan 2.6, 2.7 and 3.0: also closed, hosted only. Wan 3.0 entered public beta in August 2026.
- Newest Wan you can download: Wan 2.2 (28 July 2025, Apache 2.0), plus Wan 2.2-based Wan-Animate-2 and Wan-Dancer-14B.
- Run it locally: Wan 2.2 runs in ComfyUI on 8-16 GB GPUs with GGUF or FP8 files and 4-step LoRAs.
- Use Wan 2.5+ online: through Alibaba Cloud Model Studio’s API or third-party sites that resell it, usually on credits.
- Red flag: any “Wan 2.5/2.7 download” link is fake or a renamed Wan 2.2. Only trust the Wan-AI organisation on Hugging Face.
Is Wan 2.5 Open Source? The Short Answer
No. Alibaba’s Tongyi Lab released Wan 2.1 (February 2025) and Wan 2.2 (July 2025) as open weights under Apache 2.0. Starting with Wan 2.5-Preview, the main Wan video models have been hosted only.
As of early November 2026:
- Hugging Face: the official Wan-AI organisation has no Wan 2.5, 2.6, 2.7 or 3.0 repositories. A search for “Wan2.7” returns nothing.
- GitHub: issues asking whether 2.5 and later would be open-sourced have gone unanswered since September 2025.
- Alibaba: has not announced a policy change either way. Launch posts for 2.7 and 3.0 do not mention weights or licences.
Every Wan Version: Open or Closed
| Version | Released | Weights | Main features |
|---|---|---|---|
| Wan 2.1 | Feb 2025 | Open, Apache 2.0 | T2V and I2V at 480p/720p, 1.3B and 14B |
| Wan 2.2 | 28 Jul 2025 | Open, Apache 2.0 | MoE A14B T2V/I2V, TI2V-5B, silent ~5 s clips at 480p/720p |
| Wan 2.5-Preview | 24 Sep 2025 | Closed | Video with generated audio, higher resolution and longer clips |
| Wan 2.6 | 16 Dec 2025 (announced) | Closed | Hosted upgrade of 2.5 |
| Wan 2.7 | Apr 2026 | Closed | Four-model suite: 2-15 s clips, 720p/1080p, reference-to-video with voice, video editing, image models |
| Wan 3.0 | Beta early Aug 2026, wider launch late Aug | Closed | 2-30 s clips at 480p/720p/1080p, 30 fps, audio by default, up to 10 references |
| Wan-Animate-2, Wan-Dancer-14B | Jul-Aug 2026 | Open, Apache 2.0 | Character animation and music-to-dance, built on Wan 2.2 |
Note: the two 2026 open releases are specialist tools on the Wan 2.2 base, not general text-to-video successors.
What Wan 2.5 Added Over Wan 2.2
The jump from 2.2 to 2.5 is why people want it so badly:
- Sound: 2.5 generates synchronised audio (voices, effects, music). Wan 2.2 is silent.
- Length: longer clips than Wan 2.2’s roughly 5 seconds.
- Resolution: 1080p output, while Wan 2.2 tops out at 720p.
- Prompt following: better camera control and multi-shot scenes.
Later versions push further: Wan 2.7 added reference-to-video with a reference voice and a dedicated video-edit model, and Wan 3.0 reaches 30-second clips.
How to Use Wan 2.5 and Newer Online
Option 1: Alibaba Cloud Model Studio (official API)
- Create an Alibaba Cloud account and open Model Studio (also called Bailian / DashScope).
- Activate the video models in your region and create an API key.
- Pick a model ID, for example
wan2.5-t2v-preview,wan2.5-i2v-preview, thewan2.7-family, orwan3.0-video/wan3.0-video-prime. - Submit a job with your prompt (and an image URL for I2V), then poll for the finished video. Video calls are asynchronous.
Wan 3.0 launched at about $0.05, $0.10 and $0.20 per second for 480p, 720p and 1080p, according to providers that resell it. Check Model Studio’s pricing page for your region before you run long jobs.
Option 2: Third-party sites
Many AI video sites offer Wan 2.5-3.0 on credits next to Kling, Veo and Seedance. They are convenient, but prices vary widely and some use misleading “Wan” branding. Compare the per-second cost with Alibaba’s own rate.
For a hosted alternative with a free plan, see my Kling AI guide.
Is Wan 2.5 Free?
Not really. There is no official unlimited free tier for Wan 2.5 or newer. Your options:
- Trial credits: Alibaba Cloud and some resellers give new accounts a small free allowance.
- Reseller free plans: a few daily credits, usually with watermarks and low resolution.
- Truly free: run Wan 2.2 locally. After the GPU, every clip costs only electricity.
How to Spot Fake Wan Downloads
- Check the owner: real weights live under
Wan-AIon Hugging Face andWan-Videoon GitHub. - Distrust “Apache 2.0 Wan 2.7” claims: some SEO sites say this; it is false.
- Inspect file names: “Wan2.5” checkpoints from random uploaders are usually renamed Wan 2.2 or a fine-tune.
- Never run .exe installers that promise Wan 2.5 “offline”. ComfyUI uses
.safetensorsor.gguffiles only.
Warning: fake model downloads are a common way to spread malware. Prefer .safetensors, avoid pickled .ckpt/.pt files from unknown sources, and keep ComfyUI up to date.
Run Wan 2.2 Locally Instead (Step by Step)
Wan 2.2 is still an excellent local video model, especially with 4-step speed LoRAs. Here is the short version; my full Wan 2.2 ComfyUI guide for 8-16 GB has every workflow.
- Install or update ComfyUI (desktop or portable). RTX 50 owners should follow my RTX 50 setup guide first.
- Open the template: Workflow, Browse Templates, Video, then “Wan 2.2 14B Image to Video” or “Text to Video”.
- Download the files it asks for (table below) into the exact folders shown.
- Low VRAM? Install ComfyUI-GGUF and swap the diffusion models for Q4_K_M or Q5_K_M GGUF files.
- Add speed LoRAs (Lightning / LightX2V 4-step) to cut each clip from 20+ steps to 4.
- Generate at 480p first (832×480, 81 frames), then upscale or interpolate.
Wan 2.2 Files and VRAM Table
| File | Folder | What it is |
|---|---|---|
wan2.2_i2v_high_noise_14B_fp8_scaled.safetensors |
ComfyUI/models/diffusion_models/ |
I2V high-noise expert |
wan2.2_i2v_low_noise_14B_fp8_scaled.safetensors |
ComfyUI/models/diffusion_models/ |
I2V low-noise expert |
wan2.2_t2v_high_noise_14B_fp8_scaled.safetensors (+ low_noise) |
ComfyUI/models/diffusion_models/ |
T2V experts |
wan2.2_ti2v_5B_fp16.safetensors |
ComfyUI/models/diffusion_models/ |
Smaller 5B text+image model |
umt5_xxl_fp8_e4m3fn_scaled.safetensors |
ComfyUI/models/text_encoders/ |
Text encoder |
wan_2.1_vae.safetensors |
ComfyUI/models/vae/ |
VAE for the 14B models |
wan2.2_vae.safetensors |
ComfyUI/models/vae/ |
VAE for the 5B model |
| Your GPU | What to run | Expect |
|---|---|---|
| 8 GB | TI2V-5B or 14B Q3/Q4 GGUF, 480p | Works with offloading, slow |
| 12 GB | 14B Q4_K_M GGUF + 4-step LoRA | 480p comfortably |
| 16 GB | 14B FP8 or Q5/Q6 GGUF + 4-step LoRA | 480p fast, 720p possible |
| 24 GB+ | 14B FP8, 720p | Best local quality |
Tip: keep at least 32 GB of system RAM (64 GB is better) because ComfyUI swaps the two 14B experts in and out. My RAM for local AI guide explains why.
Open Alternatives With Audio and Longer Clips
If the features you want from Wan 2.5 are sound and longer clips, these downloadable models get closer than Wan 2.2:
| Model | Audio | Runs on 16 GB? | Guide |
|---|---|---|---|
| LTX-2.5 | Yes, joint audio-video | Yes, with GGUF | LTX-2.5 on 16 GB |
| MiniMax H3 | Yes (FastH3 Omni variant) | Yes, with GGUF/8-step LoRAs | MiniMax H3 hub (check its licence for your country) |
| Wan 2.2 + separate audio model | Add-on | Yes | Local SFX models |
For a ranked list, see the best local AI video models.
Troubleshooting
“Wan 2.5” model not found in ComfyUI Manager
Expected. There is no official local Wan 2.5. Use the Wan 2.2 templates, or Alibaba’s API through an API node.
Out of memory with Wan 2.2 14B
Drop to 480p, switch to a Q4_K_M GGUF, lower frame count to 49, and launch ComfyUI with --lowvram. My CUDA out of memory fix lists more options.
Colours look wrong or noisy output
You likely mixed VAEs. The 14B models use wan_2.1_vae.safetensors; only the 5B model uses wan2.2_vae.safetensors.
Model Studio says the model is not available
Wan models are region-specific. Switch the console region (for example Singapore for international accounts) and activate the model there.
Wan 2.2 takes 20+ minutes per clip
Add the 4-step Lightning/LightX2V LoRAs, set steps to 4 and CFG to 1, and install SageAttention. See my best Wan 2.2 LoRAs.
My Take
I would love Wan 2.5 weights too, but after more than a year of silence I would not wait for them. Alibaba now treats its flagship video models as a cloud product.
The practical answer: run Wan 2.2 with 4-step LoRAs for silent clips, LTX-2.5 when you need sound, and rent Wan 3.0 by the second only when a client job needs 1080p with audio.
FAQ
Is Wan 2.5 open source?
No. Wan 2.5-Preview was released in September 2025 as a hosted model only, with no weights or licence.
Can I run Wan 2.5 locally?
No. The newest Wan you can run locally is Wan 2.2, which works in ComfyUI on 8-16 GB GPUs with GGUF files.
Is Wan 2.7 or Wan 3.0 open source?
No. Both are closed and available only through Alibaba Cloud Model Studio and resellers.
Is Wan 2.5 free?
Only through small trial credits. Wan 2.2 is the free option if you run it on your own GPU.
Will Alibaba open-source Wan 2.5 later?
There is no announcement or date. GitHub questions about it have gone unanswered since September 2025.
What is the best open-source alternative to Wan 2.5?
LTX-2.5 for video with audio, and Wan 2.2 with speed LoRAs for silent clips.
How do I know a Wan download is real?
Real files come only from the Wan-AI organisation on Hugging Face or the Wan-Video GitHub. Anything labelled Wan 2.5 or later for download is not official.
