27B total / 14B active MoE text-to-video model with SNR-governed dynamic expert routing. High-noise expert shapes macro composition; low-noise expert refines high-frequency detail.
A workable 27B-parameter MoE video generator from Alibaba. Treat the modality benchmarks above as the leading indicator of fit — composite scoring across modalities is still maturing.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
See which devices can run this model and at what quality level.
Wan2.2-T2V-A14B is a text-to-video generation model from Alibaba, part of their Wan2.2 family. At 27B total parameters with 14B active in a Mixture-of-Experts (MoE) architecture, it occupies a unique position: it's the first open-source video generation model to use MoE routing, and it directly competes with closed-source platforms like Runway Gen-3 and Pika on output quality while remaining fully local and Apache 2.0 licensed.
The model generates 5-second videos at both 480P and 720P resolution from text prompts. What matters for practitioners is that the MoE architecture delivers better results per compute unit than dense models of similar size. Alibaba's own Wan-Bench 2.0 benchmarks show it surpassing leading commercial models across most evaluation dimensions — motion coherence, semantic alignment, and aesthetic quality.
This isn't a toy. It's a production-capable video generation model that runs on consumer hardware with appropriate quantization, and it's the strongest open-source option available for local text-to-video workloads as of mid-2025.
Wan2.2-T2V-A14B uses a Mixture-of-Experts architecture with 27B total parameters, of which 14B are active during any single forward pass. The key innovation is SNR-governed dynamic expert routing: the model separates the denoising process across timesteps using specialized expert sub-networks.
The high-noise expert handles macro composition — the broad layout, subject placement, and scene structure during early denoising steps. The low-noise expert refines high-frequency detail — textures, fine edges, and temporal consistency — during later steps. This division of labor means each expert can specialize rather than a single network trying to handle both coarse and fine-grained generation simultaneously.
For inference, this translates directly to practical advantages. You get the output quality of a 27B model while only activating 14B parameters per step, reducing VRAM pressure and inference latency compared to a dense 27B model. The MoE architecture also enables better scaling: adding more experts increases model capacity without linearly increasing per-step compute.
The model uses Wan2.2's custom VAE with a 16×16×4 compression ratio, which is aggressive by video generation standards. This means the latent space is compact, reducing the memory footprint during diffusion sampling and enabling higher resolution output without proportionally increasing VRAM requirements.
Wan2.2-T2V-A14B takes text prompts and generates video directly — no image input required. The model supports cinematic style control through detailed prompt engineering, with the training data including labeled examples of lighting, composition, contrast, and color tone.
Concrete use cases:
The model does not accept image or video inputs — it's text-only. For image-to-video workflows, Alibaba offers the I2V-A14B variant. The context window for text prompts is not officially specified, but practical usage suggests prompts of 50-200 tokens work well for coherent generation.
This is where the model's architecture pays off. The 14B active parameter count makes it feasible on consumer hardware, though you'll need to be strategic about quantization and memory management.
VRAM requirements by quantization:
Hardware recommendations:
Expected performance:
At Q4_K_M on an RTX 4090, expect approximately 3-5 tokens per second in the diffusion model, with total generation time dominated by the number of sampling steps. 50-step sampling at 720P takes roughly 3-5 minutes. Lower step counts (20-30) reduce quality but cut generation time to 1-2 minutes.
Getting started:
The quickest path is ComfyUI with the Wan2.2 integration, which handles model loading, quantization, and workflow management. The official Wan2.2 GitHub repository provides multi-GPU inference scripts for higher-end setups. Diffusers integration is also available for Python-native workflows.
vs. Stable Video Diffusion (SVD): SVD is a dense 1.4B model that generates 14-25 frame clips at 576x1024. Wan2.2-T2V-A14B produces significantly better motion coherence, higher resolution, and longer clips. The tradeoff is hardware requirements: SVD runs on 8GB VRAM, Wan2.2 needs 16GB+ for practical use. Choose SVD for quick prototyping on limited hardware; choose Wan2.2 for production-quality output.
vs. CogVideoX-5B: CogVideoX uses a dense 5B transformer architecture. Wan2.2's MoE approach with 14B active parameters delivers noticeably better composition and detail at similar inference costs. CogVideoX has a longer context window for prompts and supports image-to-video natively. Wan2.2 wins on output quality; CogVideoX wins on prompt flexibility and lower minimum VRAM (~12GB vs ~16GB).
vs. closed-source platforms (Runway Gen-3, Pika): Wan2.2 matches or exceeds these on objective quality benchmarks while running entirely locally. The tradeoffs are convenience (no API, no UI) and generation speed (slower than cloud inference). For practitioners who need privacy, unlimited generation, or production pipelines that can't depend on external APIs, Wan2.2 is the strongest open-source alternative available.
View on Hugging Face