Alibaba (Wan, Tongyi Wanxiang)

Wan (also written Wan2 or Wan2.5, and known internally at Alibaba as Tongyi Wanxiang) is the video model side of Alibaba's open-source AI strategy. Where Qwen-Image is their open-source image model, Wan is their open-source video model, and the same logic applies: release the weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable., build community, capture economic value through cloud infrastructure rather than through model gatekeeping.

Wan is the most important open-source video model in existence as of 2026. Almost any custom workflow, any community fine-tune, any ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. pipeline that involves video generation is built on Wan. Krea and Flora both offer Wan as one of their video options. The model has had a major influence on what is possible in the open-source video space, comparable to the influence FLUX has had on the open-source image space.

the distillationDistillationA technique where a smaller, faster 'student' model is trained to mimic the outputs of a larger 'teacher' model in fewer sampling steps. Used to produce fast variants of slow models (FLUX schnell, SDXL Turbo, etc.). and Mixture of Experts chapter introduced the general Mixture of Experts idea and the hospital-with-specialists analogy for how routing works. This section describes Wan's specific MoEMoE (Mixture of Experts)An architecture where multiple specialized subnetworks ('experts') are routed to handle different inputs, allowing larger total parameter counts without proportional compute increases. Used by Wan 2.2 and several language models. implementation, which is the template that other open-source video labs are now copying. If MoE is still a new concept to you, read the distillation and Mixture of Experts chapter first.

Wan 2.2, released in mid-2025, was the first open-source video generation model to use a Mixture of Experts architecture in production.

Wan 2.2 uses two expert subnetworks, both diffusion transformers, each with about 14 billion parameters. The experts are routed by noise level: a 'high-noise expert' handles the early steps of denoisingDenoisingThe core loop of diffusion. Starting from pure noise, the model repeatedly predicts and subtracts a little noise until a coherent image emerges. (where the model is establishing coarse layout and structure) and a 'low-noise expert' handles the later steps (where the model is refining details and textures). Total parameterParameterOne individual weight. Model size is quoted in parameters, in the millions or billions. More parameters can mean more capability and more cost. count is 27 billion, but only 14 billion are active per inferenceInferenceRunning a trained model to produce output. The fast, cheap part of the model lifecycle, in contrast to training. step, so the compute cost is the same as a dense 14-billion-parameter model.

This is a significant architectural insight. The reasoning is that the early and late phases of denoising are actually different problems, at high noise, the model is doing something more like layout planning, and at low noise it is doing something more like texture rendering. A single dense network has to learn both, and there is interference between the two skills. By using separate expert networks for the two phases, you let each expert specialize, and the combined model is better than either could be alone.

Wan also has a high-compression 3D VAE3D VAEA VAE that compresses both spatial and temporal dimensions of video, used to make video diffusion tractable. that achieves 4x16x16 compression ratios, meaning videos are compressed by a factor of 4 temporally and 16 in each spatial dimension, for a total compression of 1024x. This is the most aggressive compression of any open-source video VAEVAE (Variational Autoencoder)A neural network that learns to compress images into a small numerical representation (a latent) and decompress them back. Used in latent diffusion models to make the generation process tractable., and it is what makes Wan economical to run on consumer hardware.

Wan 2.1 (early 2025)

The first major release. Standard dense diffusion transformerTransformerThe neural network architecture built on attention that underpins modern language and, increasingly, image and video models. It scales well and handles long-range relationships better than convolutional networks., no MoEMoE (Mixture of Experts)An architecture where multiple specialized subnetworks ('experts') are routed to handle different inputs, allowing larger total parameter counts without proportional compute increases. Used by Wan 2.2 and several language models. yet. Established Wan as a credible open-source video model.

Wan 2.2 (mid-2025)

The MoE breakthrough release. Multiple variants: Wan2.2-T2V-A14B for text-to-videoText-to-videoGenerating video from text alone, with no starting image., Wan2.2-I2V-A14B for image-to-videoImage-to-videoStarting from a still image and generating a video that animates it according to a prompt., and Wan2.2-TI2V-5B as a smaller unified model that supports both modes and runs on consumer GPUs (a single RTX 4090 can generate a 5-second 720p clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. in a few minutes). All variants are released under the Apache 2.0 license, the most permissive open-source license, which means commercial use is fully allowed.

Wan 2.5 (current)

The flagship through early 2026. Builds on the Wan 2.2 foundation with improved architecture, better motion quality, and the addition of native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward. generation. Generates 1080p HD 10-second videos with synchronized vocals, sound effects, and music. Maintains the Apache 2.0 license. Wan 2.5 is the most capable open-source video model in existence and is roughly comparable to mid-tier closed-source models in quality, while running on consumer hardware and being completely free to use.

Wan's importance for the field is enormous. It is the model that proves you can build a high-quality video generation system entirely in the open. Every video model researcher in the world looks at what Wan is doing for techniques they can borrow. Every ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. user has a Wan workflow saved somewhere. The fact that Alibaba continues to release these models openly, despite their commercial value, is one of the most important contributions to the field's overall progress.

Wan earned its reputation as the credible open answer to Sora and Veo, and that reputation is real but increasingly out of date. As of mid-2026 the openly released weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. on Alibaba's own Hugging Face org stop at the Wan 2.2 generation, published under the permissive Apache 2.0 license along with the 1.3B and 14B sizes, the A14B Mixture-of-Experts models, and the consumer-friendly 5B. Everything newer went the other way. Wan 2.5, previewed in September 2025 with native synchronized audio, and Wan 2.6, from December 2025 with role-play character insertion and multi-shot storytelling, both shipped as closed, APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface.-only products through Alibaba Cloud with no weights published anywhere.

So the honest label for Wan today is open-core, not open. The open base models keep the community, the ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. and LoRALoRA (Low-Rank Adaptation)A fine-tuning technique that adapts a base model to a specific style or concept using a small additional file. The standard way to customize open-source models. ecosystem, and the mindshare, while the newest and most capable capabilities are held back as monetized cloud products. A Wan 2.7 appears on the mid-2026 arenaArenaA public leaderboard that ranks models from large numbers of blind pairwise human preference votes, which tends to reward average appeal over fitness for a specific job. and third-party sites claim it returns to open weightsOpen weightsModels whose trained parameters are publicly downloadable, allowing anyone to run them locally or fine-tune them. Contrasts with closed/API-only models., but the official org listed nothing past 2.2 at the time of writing, so treat the open-2.7 claim as unverified and lean toward it being closed like 2.5 and 2.6.

The strengths follow directly from the open bet. Wan is the best-in-class open-weight video model, its 1.3B and 5B variants were deliberately engineered to run on a single consumer GPUGPU (Graphics Processing Unit)The parallel processor that runs image and video models. Its speed and memory largely determine what you can generate and how fast. (a five-second 720p clipCLIPA text encoder developed by OpenAI in 2021 that learns to align text and images in a shared embedding space. Foundation of most text-to-image models from 2022 onward. in under about nine minutes on an RTX 4090), and the Apache 2.0 license lets anyone use, fine-tune, and redistribute it commercially. That combination gave it enormous ecosystem gravity: first-class ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. and Diffusers support and a deep library of community LoRAs. The 2.2 Mixture-of-Experts design also delivers 14B-class quality at closer to 14B active compute rather than paying the full expert budget every step.

The weaknesses are the cost of the open-core split. The freely downloadable ceiling is stuck at the mid-2025 2.2 generation, so the open track has fallen behind Alibaba's own closed models and the wider frontier, and native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward. exists only in the closed 2.5 and 2.6 APIAPI (Application Programming Interface)The programmatic endpoint you call to run a hosted model from your own code instead of a web interface. line. The best open models are heavy and slow unless you drop to the smaller, lower-quality variants, and the naming is genuinely confusing, with 2.1, 2.2, S2V, Animate, Dancer, 2.5, 2.6, and 2.7 all wearing the Wan badge and meaning quite different things.

The decision is really open versus hosted. For local and open-weight work, download the 2.1 and 2.2 weightsWeightsThe learned numbers inside a model that encode everything it knows. 'Open weights' means these are downloadable. from the Wan-AI org on Hugging FaceHugging FaceThe dominant repository for open-source AI models, including all the open-weight image and video models discussed in this document. or from ModelScope: pick the 1.3B or 5B for a single consumer GPUGPU (Graphics Processing Unit)The parallel processor that runs image and video models. Its speed and memory largely determine what you can generate and how fast., and use the A14B Mixture-of-Experts with more VRAMVRAM (Video RAM)The memory on a GPU. It sets a hard ceiling on the model size and image resolution you can run locally. for the best open quality, with ComfyUIComfyUIThe dominant node-based interface for running diffusion models locally. Started by 'comfyanonymous' in January 2023, now stewarded by Comfy Org. and Diffusers both fully supported. For native audioNative audioSound generated jointly with video in a single pass, so footsteps, dialogue, and ambience line up with the picture, rather than being added afterward., lip-sync, or the role-play character insertion, you need the closed 2.5 or 2.6 tier through Alibaba Cloud Model Studio or a reseller like fal or WaveSpeed. Reported hosted pricing lands around seven cents per second of video, but it is third-party and worth checking before you commit.

On the craft, write shot-style prompts with camera move, lens, lighting, and motion rather than static-image descriptions, because the models respond to cinematic direction. Use image-to-videoImage-to-videoStarting from a still image and generating a video that animates it according to a prompt., or first and last frame control on the newer hosted models, to lock composition and identity, and lean on community LoRAs to sharpen style and consistency on the open models. Keep open-model clips short, around five seconds, and reach for the hosted line when you need ten-second, 1080p, audio-synced output.