Wan is Alibaba's open-weight AI video generation model family (Wan 2.2) using a Mixture-of-Experts architecture for text-to-video and image-to-video, free to self-host with quality that rivals Kling and Runway.

Overview

Wan is an open-weight AI video generation model family developed by Alibaba’s Tongyi Lab, with Wan 2.2 as its current flagship. It supports text-to-video, image-to-video, and a combined text-and-image-to-video mode, using a Mixture-of-Experts (MoE) architecture that separates high-noise and low-noise experts to improve motion quality and prompt adherence. What sets Wan apart from Kling, Runway, or Sora is that the model weights are openly released and free to download under a permissive license — no subscription, no per-clip billing, no platform-side content filter. Since launch it has become the dominant open video model in the ComfyUI community, with extensive workflows, LoRA fine-tunes, and integrations. A smaller 5-billion-parameter TI2V variant is designed to run on a single consumer GPU while still producing usable 720p video.

Key Features

  • Mixture-of-Experts architecture — dedicated high-noise and low-noise experts improve composition and detail over single-model diffusion.
  • Text-to-video & image-to-video — plus a combined TI2V mode that blends a reference image with a text prompt for controllable output.
  • Four specialized models — T2V, I2V, character animation (Animate), and audio-driven video (S2V) cover most creator needs.
  • Consumer-grade 5B variant — the TI2V-5B model runs on ~24GB VRAM and still yields 720p@24fps clips.
  • ComfyUI-native ecosystem — first-class nodes, community workflows, and LoRA fine-tunes already available.
  • Commercially permissive license — released weights allow building products without per-generation fees.

Pricing

OptionPriceFor
Self-hosted (local GPU)$0Full open-weight access, no caps or watermarks, commercial use allowed
Cloud GPU rental$0.30–$2+/hrRun on rented instances (RunPod, Vast) without buying hardware
Alibaba hosted APIPay-per-useChinese-market billing for those who don’t want to self-host

Comparison

vs. Kling: Kling offers a polished web app and higher resolution (up to 1080p) with simpler onboarding, but bills per generation and filters content. Wan matches much of its quality for free if you can run the weights, at 720p max.

vs. Runway: Runway is the creative studio favorite with a clean UI and broad editing toolkit, ideal for non-technical users. Wan trades that polish for open weights, unlimited local generation, and fine-tuning freedom.

vs. Sora: Sora delivers some of the most coherent long-form prompts inside ChatGPT, but is closed and gated. Wan gives researchers and developers full control to modify, self-host, and build on the model.

Compare alternatives

Side-by-side with the 3 closest alternatives.

ToolCategoryPricingVisit
Wan (this) videoFrom $0/mo Site ↗
Kling AIvideoFree $0 · From $6.99/mo Site ↗
RunwayvideoFree $0 · From $15/mo Site ↗
SoravideoFrom $20/mo Site ↗
Wan Current

Wan is Alibaba's open-weight AI video generation model family (Wan 2.2) using a Mixture-of-Experts architecture for text-to-video and image-to-video, free to self-host with quality that rivals Kling and Runway.

video
From $0/mo

Kuaishou's high-quality text/image-to-video model with strong motion. Long, coherent AI video clips Read our hands-on review and compare the top AI Video

video
Free $0 · From $6.99/mo

A creator-focused AI video generation and editing platform — text-to-video and image-to-video in one place. Industry-leading generative video Read our

video
Free $0 · From $15/mo

OpenAI's text-to-video model for realistic, coherent short clips. OpenAI's most realistic video yet Read our hands-on review and compare the top AI Video

video
From $20/mo
Editor’s Review
4.6/5
Pros
  • +Fully open-weight and free to self-host — no subscription or per-generation cost like Kling or Runway
  • +MoE architecture delivers motion consistency and prompt adherence that rival closed-source leaders
  • +Lightweight 5B TI2V variant runs on a single consumer GPU (~24GB VRAM) at 720p
Cons
  • No polished consumer app — you need ComfyUI, Diffusers, or a third-party wrapper
  • Top-quality MoE models need serious hardware (40GB+ VRAM) or rented cloud GPU
  • 720p / ~10s max output; no native 1080p or 4K, and setup is technical

The most capable open-source AI video model available — Wan 2.2's MoE architecture is a real technical step forward and the 5B variant puts high-quality generation on consumer hardware. The trade-off is setup friction: this is a model, not a product, so non-technical creators should use a hosted front-end.

See all reviews →