How Alibaba Cloud’s Open‑Source Wan 2.1 Sets New Benchmarks in Video Generation

Release Overview

On 2025‑02‑25 Alibaba Cloud released the open‑source visual generation foundation model Wan 2.1. The source code, model weights and inference scripts are hosted at https://github.com/Wan-Video/Wan2.1.

Model Scale and Benchmark

Wan 2.1 is offered in a professional version with 140 billion parameters. On the VBench benchmark it achieves an overall score of 86.22 %, outperforming Sora, Luma and Pika across motion quality, visual fidelity, style consistency and multi‑objective handling.

Hardware Requirements

The 1.3 billion‑parameter variant can generate 480 p video using ≤8.2 GB VRAM, allowing execution on consumer‑grade GPUs. On an NVIDIA RTX 4090 a 5‑second 480 p clip is produced in approximately 4 minutes without quantization or additional optimizations.

Supported Tasks

Wan 2.1 provides a unified interface for:

Text‑to‑video

Image‑to‑video

Video editing (in‑place modification)

Text‑to‑image

Video‑to‑audio generation

Architecture

The model builds on the DiT (Diffusion Transformer) backbone and adopts a linear‑noise‑trajectory Flow Matching training paradigm. A causal 3D VAE encodes video frames into a latent space, and a feature‑cache mechanism enables arbitrary‑length video encoding and decoding.

Training Pipeline

Data preparation follows a four‑step cleaning pipeline: (1) collection of large‑scale image and video datasets, (2) duplicate detection, (3) quality filtering using automated metrics, and (4) final curation for diversity. Training employs distributed strategies to reduce memory consumption and increase throughput, including Fully Sharded Data Parallel (FSDP), RingAttention and the Ulysses optimizer.

Open‑Source Impact

By releasing the code and weights, Alibaba Cloud invites researchers and developers to reproduce, extend, and integrate Wan 2.1 into downstream applications, facilitating further research on multimodal video generation.