Wan2.1 — Alibaba’s Apache 2.0 Open-Source Video Model That Beat Sora on VBench
In February 2025, Alibaba released the weights to a video generation model under the Apache 2.0 license with no commercial restrictions. Within a month, the technical report showed it had achieved #1 on VBench, the industry-standard video quality benchmark, with an 86.22% total score — outperforming OpenAI’s Sora (84.28%), Runway Gen-3 (82.32%), and Tencent’s HunyuanVideo (83.24%).
The model was called Wan2.1. The name is the international shorthand for Tongyi Wanxiang (通义万相), Alibaba’s creative AI platform that also produces image generation tools under the Tongyi brand umbrella alongside Qwen LLMs. The “Wan” brand was adopted for international markets — it’s cleaner, shorter, and avoids the transliteration issues that plague Chinese AI product names abroad.
By mid-2025, Wan had released three more numbered versions, adding Mixture-of-Experts architecture, native audio, 1080P support, and eventually 4K image generation. Replicate hosts over 22 variants; wan-2.2-i2v-fast alone has accumulated 10.9 million API runs — a scale that most commercial AI video APIs never reach. The GitHub repos have accumulated over 15,000 stars each.
This is a research-based review of Wan2.1’s architecture, open-source status, capabilities, deployment, community ecosystem, and positioning relative to commercial competitors. We analyze from public sources, technical documentation, and independent benchmarks. We do not test AI tools hands-on.
The Parent Organization: Team Wan at Alibaba Cloud
Alibaba Group (阿里巴巴集团) is one of the world’s largest technology companies, founded in 1999 and headquartered in Hangzhou, China. It operates across e-commerce (Taobao, Tmall, Lazada), cloud computing (Alibaba Cloud / Aliyun), logistics (Cainiao), digital media (Youku), and AI research.
Wan is built by “Team Wan” — a dedicated group within Alibaba Cloud whose contact address is wan.ai@alibabacloud.com. The team falls under the Tongyi AI umbrella, the same brand family that produces the Qwen large language models. The technical report for Wan2.1 (arXiv:2503.20314) lists 62+ contributors from across Alibaba’s AI organization, indicating significant resource commitment. The GitHub organization is Wan-Video; the consumer product lives at wan.video (previously wanxai.com) and is mirrored at tongyi.aliyun.com/wanxiang/ for Chinese users.
It’s worth distinguishing Wan from Alibaba’s other AI video work. HappyHorse-1.0 — Alibaba’s #1-ranked anonymous launch model from April 2026 — was built by the Future Life Lab inside the Taotian Group (ATH AI Innovation Unit), not by Team Wan. Both sit within the broader Alibaba AI ecosystem, but they are organizationally and architecturally distinct. Wan is an open-source platform model; HappyHorse is a closed-weight benchmark contender with a very different architecture. The two share a corporate parent, not a development team.
The Wan2.1 technical report identifies Shiwei Zhang as the primary contact author, with the full team attributed collectively as “Team Wan, Alibaba Group.”
Version History: From 2.1 to 2.7 in Twelve Months
There is no publicly documented “Wan1.0” or “Wan2.0” open-source release. The consumer product Tongyi Wanxiang predates these releases, but the open-source program began with Wan2.1. The naming convention may reflect internal versioning or simply a decision to launch with a number that signals maturity.
| Version | Release Date | Key Additions |
|---|---|---|
| Wan2.1 | Feb 25, 2025 | T2V-14B, I2V-14B, T2V-1.3B; Apache 2.0 weights |
| Wan2.1 + ComfyUI | Feb 27, 2025 | Official ComfyUI node integration |
| Wan2.1 + Diffusers | Mar 3, 2025 | Official HuggingFace Diffusers integration |
| Wan2.1 FLF2V | Apr 17, 2025 | First-Last-Frame-to-Video (14B) |
| Wan2.1 VACE | May 14, 2025 | Video creation and editing (1.3B + 14B) |
| Wan2.2 | Jul 28, 2025 | MoE architecture (27B total / 14B active), TI2V-5B |
| Wan2.5 | ~Oct 2025 | 1080P support, 10-second video, one-pass A/V sync |
| Wan2.6 | ~Nov 2025 | Flash variants optimized for Replicate deployment |
| Wan2.7 | ~Jan 2026 | 4K image generation, up to 9 reference images, thinking mode |
The twelve-month cadence from 2.1 to 2.7 — with meaningful architectural changes at 2.1, 2.2, and 2.5 — reflects aggressive competitive pressure. Kling, Runway, and HappyHorse all released major versions in this window. Wan responded with new capabilities at each major checkpoint rather than iterating on a stable base.
No standalone arXiv papers have been published for versions beyond 2.1. Wan2.2 introduced MoE without a separate technical report; subsequent versions are described via blog posts, GitHub release notes, and community documentation.
Architecture: Diffusion Transformer with Flow Matching
Wan2.1 uses a Diffusion Transformer (DiT) with Flow Matching (rectified flows) rather than the DDPM diffusion process used in many earlier video models. Flow matching produces smoother probability paths during training, which correlates with faster inference and better visual quality at comparable model sizes.
Text Encoder: umT5
The text encoder is umT5 — a unified multilingual T5 model with 5.3 billion parameters and bidirectional attention. The choice of umT5 over alternatives (including Qwen2.5-7B, which the team considered) was made for three reasons: multilingual support without fine-tuning, bidirectional attention that reads text in context rather than left-to-right only, and faster convergence during training. umT5 embeddings are injected into each transformer block via cross-attention, allowing the video generation process to reference the full prompt at every layer rather than only at the beginning of the network.
Wan-VAE: The 3D Causal Variational Autoencoder
One of the most technically significant components of the Wan architecture is its custom VAE. Standard video generation pipelines either process videos as stacks of independent frames (losing temporal coherence) or use computationally expensive full 3D VAEs that struggle with long videos. Wan-VAE uses a 3D causal architecture with 127 million parameters and a 4×8×8 spatio-temporal compression ratio that reduces 720P video frames to a 16-channel latent space.
The key engineering achievement is a feature cache mechanism that allows encoding and decoding of 1080P videos of unlimited length by caching activations across temporal segments rather than processing the full video in memory at once. The paper reports that Wan-VAE achieves 2.5× faster inference than HunyuanVideo’s VAE, which is a meaningful practical advantage given how much inference time is consumed by encoding and decoding.
Transformer Architecture
The DiT blocks are patchified via a 3D convolution with kernel size (1,2,2) — treating video as spatiotemporal patches. The two size variants differ substantially:
- T2V-1.3B: 1536-dimensional embeddings, 8960-dimensional feedforward layers, 12 attention heads, 30 layers
- T2V-14B: 5120-dimensional embeddings, 13824-dimensional feedforward layers, 40 attention heads, 40 layers
The 14B variant is approximately 10× the parameter count of the 1.3B variant, but architectural depth increases more modestly (30→40 layers), meaning the quality improvement comes primarily from dimension scaling rather than depth.
Training: ~1 Trillion Tokens
The training pipeline follows a multi-stage curriculum:
- Low-resolution pre-training at 256px on image-text pairs
- Image-video joint training with progressive resolution scaling: 192px → 480px → 720P
- Post-training on curated high-quality video and image data
Total training volume is approximately ~1 trillion tokens. Caption quality is a critical variable in video generation training quality; Wan uses a custom captioning model built on a ViT encoder + Qwen LLM architecture that the team reports achieves caption quality comparable to Gemini 1.5 Pro. Dense, accurate captions during training are the reason Wan2.1 achieves high semantic alignment scores: the model learned from data that was described more precisely than what most training pipelines provide.
Wan2.2: Mixture-of-Experts Architecture
Wan2.2 (July 28, 2025) represents the most significant architectural change since the initial release. The update introduces Mixture-of-Experts (MoE) across the model family, with the primary T2V-A14B model having 27 billion total parameters but activating only 14 billion per inference step.
The MoE design in Wan2.2 is problem-specific rather than generic. Two specialized experts handle different phases of the denoising process:
- High-Noise Expert: Handles early denoising steps where the model is establishing overall layout, composition, and scene structure
- Low-Noise Expert: Handles late denoising steps where the model is adding fine detail, texture, and refinement
Switching between experts is triggered by a Signal-to-Noise Ratio threshold rather than by token content (as in some LLM MoE designs). This matches the actual computational problem: early and late denoising genuinely require different types of processing, and the MoE structure makes this explicit in the architecture rather than asking a single set of weights to do both.
The training data for Wan2.2 was also substantially expanded: +65.6% more images and +83.2% more videos compared to the Wan2.1 training set. The combined effect of the MoE architecture and expanded data drives the quality improvements reported in the Wan2.2 release.
The TI2V-5B variant released with Wan2.2 is notable for a different reason: it uses a new Wan2.2-VAE with a much more aggressive 16×16×4 compression ratio, which reduces memory requirements enough that the model runs on an RTX 4090 at 720P/24fps in approximately 9 minutes. This makes Wan2.2 accessible to consumer hardware for the first time — a direct response to the growing Wan community running the 1.3B model on gaming GPUs.
Performance Benchmarks
VBench: #1 at Launch
VBench is the most widely cited academic benchmark for AI video generation, evaluating multiple dimensions of video quality including subject consistency, motion smoothness, aesthetic quality, and semantic alignment with the input text.
From the Wan2.1 technical report (arXiv:2503.20314):
| Model | Quality Score | Semantic Score | Total Score |
|---|---|---|---|
| Wan 14B | 86.67% | 84.44% | 86.22% |
| Wan 1.3B | 84.92% | 80.10% | 83.96% |
| Sora (OpenAI) | 85.51% | 79.35% | 84.28% |
| MiniMax-Video-01 | 84.85% | 77.65% | 83.41% |
| HunyuanVideo | 85.09% | 75.82% | 83.24% |
| Gen-3 (Runway) | 84.11% | 75.17% | 82.32% |
| CogVideoX1.5-5B | 82.78% | 79.76% | 82.17% |
Wan 14B leads every model in this comparison on both Quality Score and Semantic Score. The 14B margin over Sora in Semantic Score (84.44% vs 79.35%) is particularly significant: semantic alignment measures how accurately the generated video reflects the input prompt, which is the dimension most directly relevant to real creative workflows. The 4B parameter gap in model size between Wan (14B) and Sora (estimated 30B+) makes this comparison even more notable.
Wan-Bench: Human Preference Evaluation
The team also conducted proprietary human preference evaluations on a benchmark called Wan-Bench, comparing to commercial models:
| Model | Weighted Score |
|---|---|
| Wan 14B | 0.724 |
| Sora | 0.700 |
| CNTopA (unnamed) | 0.693 |
| Wan 1.3B | 0.689 |
| HunyuanVideo | 0.673 |
Sub-metrics where Wan 14B leads: ID Consistency (0.946 — objects and characters remaining visually stable across frames), Physical Plausibility (0.939 — realistic motion physics), and Spatial Position Accuracy (0.590 — objects appearing where the prompt described them).
The team also reports a 69.1% win rate for Wan 14B against Runway Gen-3 in direct paired human evaluation — meaning that when human evaluators were shown one video from each model side-by-side without labels, they preferred the Wan-generated video approximately 7 out of 10 times.
Caveat: Benchmark Timing
These benchmarks were conducted as of early 2025 at Wan2.1 launch. The landscape has changed substantially. HappyHorse-1.0 (April 2026) currently holds the #1 Artificial Analysis ELO score in T2V (1,357) and I2V categories. Runway Gen-4.5 (1,247 ELO), Seedance 2.0, and Kling 3.0 have all been released since the Wan2.1 benchmarks were published. Wan’s current standing on Artificial Analysis’s live ELO leaderboard could not be confirmed — the benchmark URL returned errors during our research. Readers should verify current rankings directly at artificialanalysis.ai. The VBench #1 result remains on record and has not been retracted, but the competitive context has shifted.
What Wan Can Generate
Wan2.1 Task Suite
Text-to-Video (T2V): The baseline capability — generate video clips from a text prompt. Available at 480P and 720P in both the 1.3B and 14B variants. The 14B model produces noticeably higher quality motion and prompt alignment; the 1.3B trades quality for accessibility.
Image-to-Video (I2V): Animate a still image using a text prompt to describe the motion. 480P and 720P. A key use case is turning product photography, illustrations, or portraits into motion sequences.
First-Last-Frame-to-Video (FLF2V): Given a starting frame and an ending frame, generate the intermediate video that interpolates between them. 720P, 80 frames at 16 FPS. This capability is particularly useful for storyboard-to-animation pipelines where an artist has defined keyframes.
VACE (Video Creation and Editing): An all-in-one reference-conditioned generation, video-to-video translation, and masked video editing capability. VACE supports providing reference images or video clips to condition style, appearance, and motion. Available in 1.3B and 14B variants.
Visual Text Generation: Wan2.1 claimed to be the first video model capable of rendering readable text — both Chinese characters and English — accurately within video frames. Text-in-video has been a known failure mode of diffusion-based video models; Wan addressed it during training by including video sequences with legible text and emphasizing visual text quality metrics during evaluation.
Wan2.2 Additions
TI2V (Text+Image-to-Video): A unified model that handles both T2V and I2V within a single architecture, rather than requiring separate model weights. This simplifies deployment and enables mixed prompting: a user can provide both a text description and a reference image simultaneously. 720P at 24FPS.
S2V (Speech-to-Video): Audio-driven character animation — provide a speech audio clip and the model generates synchronized lip movements and character motion. Optional pose guidance allows controlling the character’s body position independently of the speech input. Uses CosyVoice for TTS integration in the full pipeline.
Animate: Character animation in two modes — (1) motion mimicry, where a character’s motion follows a reference video, and (2) character replacement, swapping the character’s visual identity while preserving the reference motion pattern.
Wan2.5 and Beyond
Wan2.5 (approximately October 2025) extended the platform to 1080P resolution and 10-second video duration, with one-pass audio-video sync — generating dialogue, ambient audio, and sound effects in a single inference pass rather than a separate audio pipeline step. This feature is significant: integrated audio generation has been a major differentiator for Kling 3.0 and Veo 3, and Wan2.5 brought it to the open-source ecosystem.
Wan2.7 (approximately January 2026) added 4K image generation (2K for video), support for up to 9 reference images for style and content guidance, coherent image set generation (up to 12 related images per request), and a thinking mode for enhanced compositional reasoning.
Hardware Requirements: The Consumer GPU Story
One of Wan’s most significant differentiators versus closed-source competitors is the accessibility of its weights to researchers and developers running consumer hardware.
| Model | Parameters | Minimum VRAM | Notes |
|---|---|---|---|
| T2V-1.3B | 1.3B | 8.19 GB | RTX 3070/4060 class |
| T2V-14B (with offloading) | 14B | ~24 GB | RTX 4090 / A5000 |
| I2V-14B-720P | 14B | ~24 GB | |
| FLF2V-14B | 14B | ~40 GB | A6000 / H100 class |
| VACE-1.3B | 1.3B | ~8 GB | |
| VACE-14B | 14B | ~24 GB | |
| Wan2.2 TI2V-5B | 5B | 24 GB | RTX 4090; ~9 min/clip at 720P/24fps |
| Wan2.2 T2V-A14B (MoE) | 27B total / 14B active | 80 GB+ | H100/A100 class |
| Wan2.2 I2V-A14B (MoE) | 27B total / 14B active | 80 GB+ |
The T2V-1.3B at 8.19 GB VRAM is the headline number. It means a developer or researcher with an RTX 3070 or an 8GB M1 MacBook Pro (with appropriate attention to quantization and memory-efficient attention) can run a video generation model locally with no API costs. Generation speed is slow — a 5-second 480P clip takes approximately 4 minutes on an RTX 4090, which scales proportionally worse on smaller GPUs — but the capability exists on consumer hardware.
The Wan2.2 TI2V-5B closes the quality-accessibility gap: it offers the full unified T2V/I2V capability of the 2.2 architecture at a VRAM requirement that an RTX 4090 (currently ~$1,500–$2,500 on the used market) can handle. At 9 minutes per clip, it’s a research and creative tool rather than a production pipeline, but it’s usable.
Open Source: The Apache 2.0 Advantage
Wan’s open-source model is the most commercially permissive available in AI video generation.
License: Apache 2.0 — attribution required, no restrictions on commercial use, no royalties, no restrictions on creating derivative works. This is the same license as most major open-source software projects and permits use in commercial products, APIs, fine-tuned derivatives, and integrated applications without needing to negotiate with Alibaba.
For comparison: HunyuanVideo (Tencent) uses a custom license that permits commercial use only below 100M monthly active users. CogVideoX uses a modified Apache 2.0 with additional restrictions. Mochi-1 (Genmo) uses Apache 2.0 but has a smaller model with lower quality. Wan is the most capable open-source video generation model available under a fully permissive commercial license.
GitHub Activity
| Repository | Stars | Forks |
|---|---|---|
| Wan-Video/Wan2.1 | 15,992 | 2,701 |
| Wan-Video/Wan2.2 | 15,659 | 1,933 |
| Wan-Video/Wan-skills | 41 | — |
Over 15,000 stars each for Wan2.1 and Wan2.2 within months of launch indicates strong adoption by the research and developer community — comparable to other high-profile open-source model releases.
HuggingFace Downloads
The Wan-AI organization on HuggingFace has 10,161 followers and 23 public models. Monthly download counts reflect where developer activity has concentrated:
- Wan2.2-I2V-A14B-Diffusers: 150,000/month
- Wan2.2-TI2V-5B-Diffusers: 108,000/month
- Wan2.2-T2V-A14B-Diffusers: 106,000/month
- Wan2.1-T2V-14B: 29,684/month
- Wan2.1-I2V-14B-720P: 21,740/month
- Wan2.1-T2V-1.3B: 20,927/month
The HuggingFace Spaces community demo (the web interface to try Wan without local deployment) has accumulated 2,090+ likes and 43 discussions. The download numbers suggest Wan2.2 variants are seeing an order of magnitude more adoption than the original 2.1 release — either due to quality improvements or because the community shifted to the Diffusers integration as the standard deployment path.
API Access and Third-Party Deployment
Alibaba Cloud DashScope
The enterprise and developer API is available via Alibaba Cloud DashScope (Model Studio). International users access it via dashscope-intl.aliyuncs.com. The API supports async task submission with status polling and callback URLs. Prompt enhancement uses Alibaba’s Qwen models (qwen-plus, qwen-vl-max) as a preprocessing step. Specific per-video pricing was not available in publicly accessible documentation during our research; developers should check the Alibaba Cloud pricing page directly, as API pricing is dynamically loaded and region-dependent.
Replicate: 10.9 Million Runs
Replicate hosts over 22 Wan model variants under the wan-video organization. Run counts as of our research:
| Model | Runs |
|---|---|
| wan-2.2-i2v-fast | 10,900,000 |
| wan-2.2-5b-fast | 593,300 |
| wan-2.5-i2v | 210,500 |
| wan-2.5-i2v-fast | 66,300 |
| wan2.6-i2v-flash | 66,400 |
| wan-2.2-s2v | 109,400 |
| wan-2.7-image-pro | 44,000 |
The wan-2.2-i2v-fast figure of 10.9 million runs is the headline number — it suggests that image-to-video is the primary use case driving API adoption, and the “fast” variants optimized for lower latency are preferred over the standard quality variants in API usage. This makes sense for developers integrating video generation into products where response time matters.
fal.ai and Additional Platforms
fal.ai hosts multiple Wan variants with per-second pricing (specific rates were not available in public documentation at research time — check fal.ai directly). Additional platforms include Replicate, ModelScope (Alibaba’s own model hub), and direct ComfyUI deployment. The Diffusers integration (pip install diffusers) makes Wan accessible from Python code with a standard API that requires no platform-specific SDK.
Integrations and Developer Ecosystem
ComfyUI: Official node integration released February 27, 2025 — two days after the weights. ComfyUI is the dominant workflow tool for local AI image and video generation; official support on day two indicates the team prioritized community adoption from the start.
HuggingFace Diffusers: Official integration March 3, 2025. The Diffusers library is the standard Python API for diffusion model inference; Wan being available via diffusers means it integrates with any existing Diffusers-based pipeline.
PyTorch FSDP + DeepSpeed Ulysses: Multi-GPU training parallelism is built into the Wan codebase, enabling fine-tuning on multi-GPU clusters. The architecture supports both data parallelism and sequence parallelism.
xDiT / xfuser: Universal Sequence Parallelism libraries for distributed inference, enabling multi-GPU deployment for lower latency in production environments.
CosyVoice: Alibaba’s TTS model, integrated for the S2V (Speech-to-Video) pipeline in Wan2.2 and later.
Wan-skills (GitHub: Wan-Video/Wan-skills): A collection of AI agent skill definitions for image generation via the DashScope API. This is not an MCP server in the Anthropic Model Context Protocol sense — it provides skill specifications for agent frameworks but doesn’t implement the MCP protocol. The repository has 41 GitHub stars and was last updated March 31, 2026.
MCP Server Status
There is no official MCP server for Wan, either from Team Wan or from third-party developers in the MCP ecosystem. The Wan-skills repository described above is sometimes referenced in this context, but it implements a different interface spec.
For developers wanting to use Wan capabilities via an MCP-compatible interface, the practical options are: (1) write a custom MCP wrapper around the DashScope API, (2) write a custom MCP wrapper around the Replicate API for Wan variants, or (3) wait for official MCP support, which the team has not announced.
Given the rapid release cadence and growing API adoption, an official or community MCP server for Wan is plausible within the next 1–2 major versions.
Consumer Product: wan.video
The consumer-facing product at wan.video (also accessible at create.wan.video and tongyi.aliyun.com/wanxiang/ for Chinese users) offers 40+ AI creative features including text-to-image, image-to-image, text-to-video, image-to-video, style transfer, super resolution, video repainting, sketch-to-image, and cartoon avatar generation.
The platform requires login to access. Pricing is not exposed in static HTML — it uses a credit or subscription model loaded dynamically, consistent with most AI video consumer products. An international version is confirmed (the codebase references wan-app-internation). Specific pricing tiers were not available in our research; check wan.video directly.
The consumer product appears secondary to the API and open-source strategy in the team’s communications. Public announcements and technical documentation focus on model capabilities, GitHub/HuggingFace releases, and API access. This contrasts with Kling and Pika, where the consumer app is the primary user-facing product.
Competitive Positioning
Wan occupies a specific niche in the video generation market that no other major model fills: fully open-source (Apache 2.0), commercially permissive, top-tier quality, consumer GPU accessible.
Against each major competitor:
vs. Kling 3.0: Kling is entirely closed-source. It offers native 4K, multi-shot storyboarding, and full native audio — capabilities Wan has partially addressed in 2.5/2.7 but not at parity. However, Wan can be self-hosted, fine-tuned, and deployed without paying per-clip fees. Enterprise customers with privacy requirements or high-volume workflows have no comparable alternative to Wan.
vs. Runway Gen-4: Runway’s models are closed, API-only, and focused on the professional creative market. No self-hosting option exists. Wan2.1’s 69.1% head-to-head win rate against Runway Gen-3 in human evaluation is the strongest evidence that the quality gap has been closed for general content.
vs. HunyuanVideo (Tencent): The closest structural comparison — both are open-source Chinese AI models with strong benchmarks. HunyuanVideo uses a custom license capping commercial use at 100M MAU; Wan’s Apache 2.0 license has no such restriction. Wan2.1’s VBench score (86.22%) leads HunyuanVideo (83.24%). HunyuanVideo has strong community adoption but Wan’s download numbers on HuggingFace now exceed it in the 2.2 generation.
vs. HappyHorse-1.0 (Alibaba / Taotian): Within the Alibaba ecosystem, HappyHorse currently holds the #1 Artificial Analysis ELO for T2V, with a unified audio-video architecture and 38-second 1080P generation on a single H100. HappyHorse’s weights are open-source (Apache 2.0) but require H100-class hardware, pricing out consumer GPU deployment. Wan TI2V-5B (24GB VRAM) is the accessible alternative for developers who can’t afford H100 access. The two models serve different hardware tiers within the same corporate portfolio.
vs. Sora: OpenAI’s Sora is entirely closed (discontinued April 2026 as a standalone product, per our earlier coverage). Wan2.1’s VBench score (86.22%) was already above Sora (84.28%) at launch. The comparison is now moot for practical deployment but remains relevant as a benchmark reference.
What the Research Community Has Built
The 10.9 million Replicate runs and 2,090+ HuggingFace Spaces likes represent something beyond API usage — they reflect a community that has made Wan the de facto reference implementation for video generation research.
Specific community applications visible in HuggingFace discussions and GitHub issues include: character animation for indie games, product demo video automation, architectural visualization, social content creation pipelines, and educational video generation. The VACE module has attracted particular attention for video editing workflows where users want to modify specific regions of existing videos.
ComfyUI integration is where most of the downstream tooling lives — custom nodes, workflow templates, and automation scripts that extend Wan’s capabilities for specific use cases. Because ComfyUI is itself open source and has an active plugin ecosystem, Wan benefits from a multiplier effect where community improvements to the ecosystem compound with each other.
What Wan Doesn’t Have (Yet)
No official MCP server: Builders who want to integrate Wan into Claude, Cursor, or other MCP-compatible environments need to write their own server or use a generic Replicate/DashScope wrapper.
No Adobe integration: Unlike Kling (Adobe Firefly), Pika (no, but similar tier), and others, Wan has no integration into professional creative tools. The consumer product is web-based; there’s no Premiere Pro plugin or Creative Cloud connection.
No native storyboarding: Kling 3.0’s multi-shot storyboarding (6 scenes per session with maintained character consistency) is a workflow capability that Wan doesn’t currently match. Wan supports I2V and FLF2V but not multi-scene coherent generation in a single pass.
Slow generation on consumer hardware: A 5-second 480P clip takes ~4 minutes on an RTX 4090. For production-volume video generation, this isn’t viable without multi-GPU setups or commercial API access. The fal.ai and Replicate APIs are the practical solution for production use cases.
Uncertain current benchmark standing: Wan’s position on the Artificial Analysis live ELO leaderboard was not verifiable during research. Given the wave of high-quality commercial models released since early 2025, Wan may have dropped from its VBench #1 position in real-world user preference. The benchmark results in this review are from early 2025 and may not reflect the current quality gap.
Rating: 4 out of 5
Wan is the most important open-source video generation model available. Apache 2.0 licensing, consumer GPU accessibility (8GB VRAM for the 1.3B model), VBench #1 at launch, 10.9 million Replicate runs, and a twelve-month release cadence that now reaches 4K output and one-pass audio-video sync — this is a production-ready platform, not a research demo.
The four-out-of-five is appropriate rather than five-out-of-five for three reasons. First, the ELO standings against the current competitive field (HappyHorse, Kling 3.0, Seedance 2.0) are unknown — the VBench result is a point-in-time measurement from a rapidly moving benchmark landscape. Second, the no-MCP gap is a real integration friction point for the growing ecosystem of Claude-based and MCP-compatible agent tools. Third, consumer hardware generation speed (4 minutes per 5-second clip on an RTX 4090) means that the open-source self-hosting story has real practical limitations outside of research contexts.
What Wan does better than any commercial closed-source alternative: it can be fine-tuned, self-hosted, embedded in commercial products without per-clip fees, and studied architecturally by researchers. For the significant portion of the AI video market that needs those properties, Wan2.x is the obvious choice.
This is a research-based review by ChatForest. We analyze AI video generation tools from public sources, technical documentation, academic papers, and independent benchmarks. We do not test tools hands-on. Information is current as of May 2026; specific pricing and benchmark standings should be verified directly with providers.