Wan 2.2

Alibaba's open-weights video model family, released July 28, 2025 under Apache 2.0. A 27B mixture-of-experts diffusion transformer with 14B active parameters, plus a 5B variant that runs on an RTX 4090. The later Wan 2.5, 2.6 and 2.7 releases are closed, API-only.

Released
Jul 28, 2025
Type
Video Generation Model
License
Apache 2.0
🌐 Official Website📚 Documentation💰 Pricing & API
On this page

Overview

Wan is Alibaba's video generation family, built by Tongyi Lab and marketed in China as 通义万相 (Tongyi Wanxiang). Wan 2.2 was released on July 28, 2025, and it is the part of the family you can actually download. The model card's news log is unambiguous: "Jul 28, 2025: We've released the inference code and model weights of Wan2.2."

Its headline architectural claim is that "Wan2.2 introduces a Mixture-of-Experts (MoE) architecture into video diffusion models." The two large variants carry 27B total parameters with 14B active per denoising step, split into "a high-noise expert for the early stages, focusing on overall layout; and a low-noise expert for the later stages, refining video details." Alongside them Alibaba shipped a 5B dense model, TI2V-5B, built on a Wan2.2-VAE (autoencoder) with a 4×16×16 compression ratio, which "can generate a 5-second 720P video in under 9 minutes on a single consumer-grade GPU."

That 5B model is why Wan matters. It made competent local video generation a consumer-hardware activity, and the community built accordingly: Comfy-Org/Wan_2.2_ComfyUI_Repackaged alone recorded roughly 4.24 million Hugging Face downloads in the last 30 days, and 452 Hugging Face repositories declare Wan-AI/Wan2.2-I2V-A14B as their base model. The GitHub repository has 16,564 stars and 2,059 forks.

Alibaba has continued the Wan line — 2.5, 2.6, 2.7 — but has not open-sourced any of it since 2.2.

A note on versions and licensing

Two claims circulate widely on the web. Both are false.

"Wan 3.0" does not exist. Alibaba has never announced it. The Wan-Video GitHub organization contains exactly four repositories — Wan2.1, Wan2.2, Wan-skills, and a diffusers fork — and the newest model code is Wan2.2. A Hugging Face hub-wide search for Wan3.0 returns zero models. The blogs promoting "Wan 3.0" do not agree with each other: one advertises "Alibaba's Next-Gen 60B Open-Source AI Video Generator," while another describes the same model as having "14 billion" and "1.3 billion" parameter variants — which are Wan 2.1's configurations — and lists references ("Wan 2.1 GitHub Repository," "Hugging Face — Wan Video Models") with no URLs behind them.

Wan 2.7 is not Apache 2.0, and has no open weights. Wan 2.5-Preview was announced at Alibaba's Apsara Conference on September 24, 2025; the Wan2.6 series followed on December 16, 2025; Wan2.7-Image on April 1, 2026 and Wan2.7-Video the week after (Alibaba Cloud's post is dated April 7, 2026). Each announcement routes users to Alibaba's hosted products and nothing else: the Wan2.6 press release says users "can access and deploy the models through Model Studio—Alibaba Cloud's AI development platform—and Wan's official website," while the Wan2.7-Video post says both models are "now available on Alibaba Cloud's Model Studio and the official Wan website." Neither mentions open source, weights, or a license. There is no Wan-AI/Wan2.7-* repository on Hugging Face; a hub-wide search for "Wan2.7" returns zero models. The community noticed: GitHub issue #184 ("Is WAN 2.5 going to be open source?") and #181 ("Thank you Alibaba for deceiving and using the open source community") were both opened on September 24, 2025 and both remain open, with no maintainer answer.

So: Wan 2.2 is Apache 2.0. Wan 2.5, 2.6 and 2.7 are closed, API-only products. When a benchmark table shows "Wan" near the top, check which one it means.

Capabilities

  • Text-to-video: Wan2.2-T2V-A14B generates 5-second clips at 480P and 720P from a text prompt.
  • Image-to-video: Wan2.2-I2V-A14B animates a still image at 480P and 720P.
  • Unified text-and-image-to-video on consumer hardware: Wan2.2-TI2V-5B covers both tasks in one 5B model at 720P/24fps, on a 24GB GPU.
  • Speech-to-video: Wan2.2-S2V-14B (August 26, 2025) is "an audio-driven cinematic video generation model" — it animates a portrait to a supplied audio track.
  • Character animation and replacement: Wan2.2-Animate-14B (September 19, 2025) is "an unified model for character animation and replacement with holistic movement and expression replication."
  • Cinematic controls: Alibaba curated "aesthetic data, complete with detailed labels for lighting, composition, contrast, color tone" to make style steerable from the prompt.
  • Reference-to-video, video editing, and Thinking Mode exist only in the closed 2.6 and 2.7 API models — wan2.7-r2v-2026-06-12, wan2.7-videoedit, and the thinking_mode parameter on the Wan-Image 2.7 API. None of this is in the open weights.

Technical Specifications

  • Open checkpoints (Apache 2.0): Wan-AI/Wan2.2-T2V-A14B, Wan-AI/Wan2.2-I2V-A14B, Wan-AI/Wan2.2-TI2V-5B, Wan-AI/Wan2.2-S2V-14B, Wan-AI/Wan2.2-Animate-14B — plus -Diffusers variants of each
  • Architecture: MoE diffusion transformer; two-expert denoising split by signal-to-noise-ratio threshold
  • Parameters: 27B total, 14B active per step (A14B models); 5B dense (TI2V-5B)
  • VAE: Wan2.2-VAE, 4×16×16 compression; TI2V-5B reaches 4×32×32 with an added patchification layer
  • Resolution / frame rate: 480P and 720P (A14B); 720P at 24fps (TI2V-5B). Clip length 5 seconds
  • VRAM: "at least 80GB VRAM" for single-GPU A14B inference; "at least 24GB VRAM (e.g, RTX 4090 GPU)" for TI2V-5B
  • Training data scale: "+65.6% more images and +83.2% more videos" than Wan 2.1 (relative, not absolute — Alibaba publishes no absolute counts)
  • License: Apache 2.0. "We claim no rights over the your generated contents"
  • Technical report: arXiv:2503.20314, Wan: Open and Advanced Large-Scale Video Generative Models (submitted March 26, 2025; v2 April 19, 2025), which covers the 1.3B and 14B Wan generation

Alibaba publishes no absolute training-token or training-hour figure for Wan 2.2, so this page does not state one.

Use Cases

  • Local video generation: The dominant use. TI2V-5B on a 24GB consumer GPU, or a quantized A14B via GGUF, with no API dependency and no per-second billing.
  • ComfyUI production pipelines: Node-based workflows chaining Wan 2.2 with upscalers, interpolators, and LoRAs.
  • LoRA fine-tuning and character consistency: 452 Hugging Face repositories are built directly on Wan2.2-I2V-A14B, most of them LoRAs.
  • Talking-head and avatar video: Wan2.2-S2V-14B for audio-driven portrait animation; Wan2.2-Animate-14B for motion transfer and character replacement.
  • Research baselines: An Apache 2.0 MoE video diffusion model with a published technical report is a rare, citable starting point.
  • Hosted API work at higher quality: For teams that only need output, the closed wan2.7-* models on Model Studio are substantially better than the open ones — at the cost of the local-inference story.

Performance / Benchmarks

Alibaba's own claim, from the Wan2.2 model card, is that expanded training data yields "TOP performance among all open-sourced and closed-sourced models," and that "On our new benchmark Wan-Bench 2.0, the model surpasses leading commercial models across most key evaluation dimensions." Wan-Bench 2.0 is Alibaba's own benchmark, the claim dates from July 2025, and independent leaderboards do not support it today.

arena.ai (formerly LMArena) ranks Wan versions against each other on a single Elo scale. Snapshots: image-to-video, June 23, 2026, 1,350,288 votes; text-to-video, July 5, 2026, 533,418 votes.

ModelWeightsarena.ai image-to-videoarena.ai text-to-video
wan2.7-i2v / wan2.7-t2vClosed, API only#5 — Elo 1434 ±8#11 — Elo 1348 ±10
wan2.6-i2v / wan2.6-t2vClosed, API only#17 — Elo 1316 ±10#15 — Elo 1333 ±10
wan-v2.2-a14bApache 2.0#37 — Elo 1169 ±10#36 — Elo 1131 ±15

Artificial Analysis — a different organization with a different Elo scale, which must never be cross-compared with arena.ai — places Wan2.7-260612 (the wan2.7-t2v-2026-06-12 snapshot) at #2 on its text-to-video "With Audio" leaderboard with an Elo of 1160, behind Dreamina Seedance 2.0 720p at 1224. On its image-to-video "With Audio" leaderboard, Wan 2.7 ranks #4 with an Elo of 1096. Artificial Analysis splits each video board into a "With Audio" and a "No Audio" view; because the open Wan 2.2 generates silent video, it never appears alongside Wan 2.7 on the With Audio boards.

Every strong Wan result on either site belongs to a model you cannot download.

The open-weights video gap

There is a real asymmetry between open image models and open video models, and Wan sits on the wrong side of it.

On Artificial Analysis's text-to-image arena, the leading open-weights entry is Cosmos3-Super-Text2Image at Elo 1226 (NVIDIA), followed by HiDream-O1-Image-Dev-2604 (1187) and ERNIE Image (1168) — two of the top three from Chinese labs. Its video boards are harder to read, because they split by audio. On the "With Audio" boards the only open-weights entries are Lightricks' LTX family — LTX-2.3 Fast at #24 (Elo 977) and LTX-2.3 Pro at #25 (962) on text-to-video, LTX-2.3 Pro at #21 (952) on image-to-video — and silent Wan 2.2 cannot compete there at all. On the "No Audio" text-to-video scale, Wan 2.2 A14B does appear, at Elo 1114: ahead of LTX-2.3 Pro (1108) and behind LTX-2.3 Fast (1127). Lightricks leads open-weights video, but it has not lapped Wan 2.2.

Wan 2.2 still owns the ecosystem — the ComfyUI workflows, the GGUFs, the LoRAs, the distillations. But it was released in July 2025, Alibaba has shipped three closed generations since, and its Elo on arena.ai now trails its own closed sibling by 217 points on text-to-video and 265 on image-to-video. The open weights have the users; the closed API has the quality.

Limitations

  • The open model is three generations old. Wan 2.2 dates from July 2025. Wan 2.5, 2.6 and 2.7 all shipped since, none with weights.
  • Vendor benchmarks are stale and self-scored. "TOP performance among all open-sourced and closed-sourced models" was Alibaba's July 2025 claim on its own Wan-Bench 2.0. Public arena results now place wan-v2.2-a14b in the mid-30s on both arena.ai video boards.
  • No audio generation. The open T2V and I2V models output silent video; Alibaba Cloud's Model Studio table lists the hosted wan2.2-t2v-plus with the feature "No audio." S2V consumes audio, it does not synthesize it.
  • Serious hardware for the good variants. Official single-GPU inference for the A14B models wants "at least 80GB VRAM." The 24GB path is the 5B model or a community quantization.
  • Short clips. 5 seconds, 480P/720P. The closed 2.7 API reaches 2–15 seconds at 1080P.
  • The open-source commitment is not a commitment. Two GitHub issues asking whether Wan 2.5+ would be opened have sat unanswered since September 24, 2025. Plan as though 2.2 is the last open release, because today it is.
  • Apache 2.0 covers the weights only. Training data, the data pipeline, and Wan-Bench 2.0's contents are not published.

Pricing & Access

Self-hosting (Wan 2.2)

Free, under Apache 2.0, with no revenue cap and no field-of-use restriction. Weights are at Wan-AI on Hugging Face and ModelScope; inference code at Wan-Video/Wan2.2.

Hosted (Wan 2.2 and the closed models)

Alibaba Cloud Model Studio serves the whole family. Verified model IDs include wan2.7-t2v-2026-06-12, wan2.7-i2v-2026-04-25, wan2.7-r2v-2026-06-12, wan2.7-videoedit, wan2.7-image, wan2.7-image-pro; the wan2.6-* and wan2.5-*-preview families; and wan2.2-t2v-plus, wan2.2-i2v-plus, wan2.2-animate-move, wan2.2-animate-mix. The older wan2.1-* models are still listed and still priced.

Alibaba Cloud bills video output by duration — "Cost = Video unit price × Video duration (seconds)" — and notes that "Some models charge by output video resolution." Its published rate card (last updated July 8, 2026):

Model IDResolutionFirst-party price
wan2.7-t2v / wan2.7-i2v720P$0.10 / second
wan2.7-t2v / wan2.7-i2v1080P$0.15 / second
wan2.6-t2v720P / 1080P$0.10 / $0.15 per second
wan2.2-t2v-plus / wan2.2-i2v-plus480P$0.02 / second
wan2.2-t2v-plus / wan2.2-i2v-plus1080P$0.10 / second

Third-party figures, attributed:

SourceModelPublished price
Artificial AnalysisWan 2.7$9.00 per minute of output
Artificial AnalysisWan 2.6$9.00 per minute of output
Together AI pricing pageWan 2.2 I2V$0.31 per video
Together AI pricing pageWan 2.2 T2V$0.66 per video

Consumer access to the closed models runs through wan.video and, per Alibaba, integration into the Qwen App.

Ecosystem & Tools

  • ComfyUI — official native workflows for TI2V-5B, T2V-A14B and I2V-A14B; the 5B "should fit well on 8GB vram with the ComfyUI native offloading"
  • Comfy-Org/Wan_2.2_ComfyUI_Repackaged — ~4.24M downloads in 30 days, the most-used Wan artifact on Hugging Face
  • GGUF quantizations — QuantStack/Wan2.2-I2V-A14B-GGUF (~184k/30d) and QuantStack/Wan2.2-T2V-A14B-GGUF (~83k/30d); more than 50 Hugging Face repositories carry both "Wan2.2" and "GGUF" in the name
  • Distillations and speed LoRAs — lightx2v/Wan2.2-Distill-Loras (~294k/30d)
  • Diffusers — first-party -Diffusers checkpoints for every Wan 2.2 model, including Animate-14B as of November 13, 2025
  • Wan-Video/Wan-skills — Apache 2.0 agent skills that call the closed Wan APIs; requires an Alibaba Cloud account and a Model Studio API key
  • Hosted inference — Together AI, and Alibaba Cloud Model Studio for the whole family

Community & Resources

Frequently Asked Questions

July 28, 2025. The Wan2.2 model card's news log reads: "Jul 28, 2025: We've released the inference code and model weights of Wan2.2." Alibaba Cloud's press release followed on July 29, 2025. The Wan-Video/Wan2.2 GitHub repository was created on July 28, 2025.
Yes. The GitHub README states: "The models in this repository are licensed under the Apache 2.0 License." Every Wan-AI checkpoint on Hugging Face carries the apache-2.0 license tag. Alibaba adds: "We claim no rights over the your generated contents." Training data is not released.
No. This is a widely repeated falsehood. Wan 2.7 exists only as an API on Alibaba Cloud Model Studio and wan.video. There is no Wan-AI/Wan2.7-* repository on Hugging Face — a hub-wide search for "Wan2.7" returns zero models — and no Wan2.7 repository in the Wan-Video GitHub organization. Alibaba's own Wan2.7 announcements say nothing about open sourcing, weights, or a license.
No. Alibaba has never announced one. The newest repository in the Wan-Video GitHub organization is Wan2.2, and the newest weights in the Wan-AI Hugging Face organization are Wan2.2. Pages describing "Wan 3.0" contradict each other — one claims 60 billion parameters, another recycles Wan 2.1's 14B and 1.3B figures — and none link to an Alibaba source.
Only Wan 2.1 and Wan 2.2. The Wan-AI Hugging Face organization holds 23 models, all of them Wan2.1 or Wan2.2. Wan 2.5, 2.6 and 2.7 have never shipped checkpoints.
The official inference script for the A14B mixture-of-experts models requires "a GPU with at least 80GB VRAM." The 5B TI2V model "can run on a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)," and ComfyUI's guide says the 5B "should fit well on 8GB vram with the ComfyUI native offloading." Community GGUF quantizations lower these figures further.
The text-to-video and image-to-video models produce silent video — Alibaba Cloud's Model Studio table lists the hosted wan2.2-t2v-plus and wan2.2-i2v-plus with the feature "No audio." The separate Wan2.2-S2V-14B is audio-driven: it consumes an audio track and animates to it, rather than generating sound.
Self-hosting Wan 2.2 is free under Apache 2.0. For hosted access, Alibaba Cloud bills output by duration — "Cost = Video unit price × Video duration (seconds)" — and charges by resolution. Its published rate card lists wan2.7-t2v and wan2.7-i2v at $0.10/second at 720P and $0.15/second at 1080P, and wan2.2-t2v-plus/wan2.2-i2v-plus at $0.02/second at 480P and $0.10/second at 1080P. Third parties publish their own figures: Artificial Analysis lists Wan 2.7 at $9.00 per minute of output, and Together AI lists Wan 2.2 I2V at $0.31 per video and Wan 2.2 T2V at $0.66 per video.
Far behind. On arena.ai's image-to-video leaderboard the closed wan2.7-i2v sits at rank 5 with an Elo of 1434, while the open wan-v2.2-a14b sits at rank 37 with an Elo of 1169. On the text-to-video board it is rank 11 (1348) versus rank 36 (1131). Same leaderboard, same Elo scale: a gap of 265 points on image-to-video and 217 points on text-to-video.
A planning step in the newer closed models. Alibaba's Wan-Image 2.7 API reference documents a thinking_mode boolean that "Defaults to true" and, when enabled, means "the model enhances its inference capabilities to improve output quality, but this increases generation time." It is documented for the image models; Alibaba's Wan 2.7 image-to-video API reference does not mention it.

Related Models

Sora 2

OpenAI's video generation model with synced audio, released September 2025. The consumer app shut down in April 2026; the API remains live until September 24, 2026.

Kling 3.0

Kuaishou's closed-weight video generation family, launched February 5, 2026. Four SKUs (Video 3.0, Video 3.0 Omni, Image 3.0, Image 3.0 Omni), clips of 3–15 seconds with natively generated audio, and a business doing roughly USD 500 million in annualized revenue.

Seedance 2.0

ByteDance Seed's video generation model, launched in China on February 12, 2026. A unified multimodal audio-video joint generation architecture taking text, image, audio and video inputs; its global rollout was paused in March 2026 after cease-and-desist letters from Hollywood studios.

Hailuo 2.3

MiniMax's flagship video generation model, released October 28, 2025. Closed weights, text-to-video and image-to-video up to 1080p, no audio track — and still MiniMax's newest video model more than eight months later, while nine other MiniMax releases shipped in the interim.

HappyHorse 1.1

Alibaba's video generation model, released June 23, 2026. Built by the Alibaba Token Hub (ATH) business unit rather than the Tongyi team behind Wan, it generates up to 15 seconds of 1080p video with synchronized audio, and HappyHorse-1.0 still leads Artificial Analysis's no-audio text-to-video arena.

Vidu Q3 Pro

ShengShu Technology's flagship video generation model, announced January 30, 2026. Generates up to 16 seconds of 1080p video with native audio in a single pass. Reference-to-video runs on sibling Q3 model IDs, not on Q3 Pro itself. ShengShu shipped a separate real-time model, Vidu S1, on July 3, 2026.

Explore More Models

Discover other AI models and compare their capabilities.

Browse Models← Back to Models