T2V-A14B MoE design
According to the model card, the A14B series uses two denoising experts for high-noise layout and low-noise detail, each about 14B parameters for 27B total with 14B active per step.
Open Source Model Profile · Wan-AI
Wan2.2-T2V-A14B
Wan2.2-T2V-A14B is a text-to-video model from Wan-AI. According to the model card, this repository holds the T2V-A14B Mixture-of-Experts variant supporting 5-second videos at 480P and 720P.
- Publisher
- Wan-AI
- Task
- text-to-video
- Model type
- Unknown
- License
- apache-2.0
- Library
- wan2.2
- Publication status
- Accepted · not indexed
Model overview
Wan2.2-T2V-A14B is published by Wan-AI as a text-to-video checkpoint using the wan2.2 library. Hub tags record diffusers, safetensors, and Apache-2.0 licensing. According to the model card, Wan2.2 is a major upgrade introducing Mixture-of-Experts video diffusion with cinematic-aesthetics curation and expanded motion training data.
Recorded capabilities
5-second 480P and 720P output
According to the model card, this repository supports generating 5-second videos at both 480P and 720P resolutions.
Documented GPU workflows
According to the model card, a single-GPU command can run on a GPU with at least 80GB of VRAM, with offload and dtype flags available for memory errors and FSDP plus Ulysses for multi-GPU runs.
Prompt-extension options
According to the model card, prompt extension through the Dashscope API or a local Qwen model can enrich details and improve video quality.
Use cases in the source record
- Text-to-video generation of 5-second clips at 480P and 720P using the documented generate.py task flag for t2v-A14B.
- Single-GPU inference with memory-saving flags such as model offloading, dtype conversion, and CPU text-encoder placement when encountering out-of-memory errors.
- Multi-GPU inference with PyTorch FSDP and DeepSpeed Ulysses acceleration, plus optional Dashscope or local-model prompt extension for richer video detail.
Limitations and unknowns
- According to the model card, top-performance and benchmark-superiority statements are publisher claims on Wan-Bench 2.0; no independent evaluation results were extracted.
- Consumer-GPU and speed statements about a separate 5B TI2V variant do not describe this A14B repository and should not be read as its requirements.
- No Safetensors parameter count, architecture identifiers, or context-window value was extracted for this repository.
- Provider state is historical snapshot data, not independently refreshed current availability.
Source and provenance
Source: Wan-AI/Wan2.2-T2V-A14B
Captured: Unknown. Processed: 2026-09-07T19:34:38.557955+00:00.
Wan2.2 💜 Wan ｜ 🖥️ GitHub | 🤗 Hugging Face | 🤖 ModelScope | 📑 Technical Report | 📑 Blog | 💬 WeChat Group | 📖 Discord Wan: Open and Advanced Large-Scale Video Generative Models We are excited to introduce Wan2.2 , a major upgrade to our foundational video models. With Wan2.2 , we have focused on incorporating the following innovations: 👍 Effective MoE Architecture : Wan2.2 introduces a Mixture-of-Experts (MoE) architecture into video diffusion models. By separating the denoising process cross timesteps with specialized powerful expert models, this enlarges the overall model capacity while maintaining the same computational co…
F001F002F003F004F005F006F007F008F009F010F011F012F013F016F017F018F019F020F021F026F028F030