Alibaba Cloud's Wanxiang 2.1: Open‑Source Dual‑Version Visual Generation Model with Full‑Scale Capabilities

Alibaba Cloud announced the open‑source release of Wanxiang 2.1, a visual generation foundation model that comes in two parameter scales: a 140 billion‑parameter professional version for high‑quality output and a 13 billion‑parameter version that runs on consumer‑grade GPUs with fast inference.

The models achieve state‑of‑the‑art results, surpassing existing open‑source and commercial solutions on benchmarks such as VBench, where the 140 B model scores 86.22% and outperforms Sora, Luma, and Pika. The 13 B model generates 480p video using only 8.2 GB VRAM, completing a 5‑second clip on an RTX 4090 in about four minutes without quantization.

Wanxiang 2.1 supports multiple generation tasks, including text‑to‑video, image‑to‑video, video editing, text‑to‑image, and audio synthesis, and is the first open‑source model capable of rendering both Chinese and English text directly within videos.

The model incorporates a powerful video VAE (Wan‑VAE) that can encode and decode arbitrarily long 1080p videos while preserving temporal information, and its evaluation covers 14 major dimensions and 26 sub‑dimensions, achieving top scores in five categories such as complex motion, physical realism, and visual quality.

Architecturally, Wanxiang 2.1 builds on the mainstream video DiT framework with Full‑Attention, uses a linear‑noise‑trajectory Flow Matching training paradigm, and integrates a self‑developed causal 3D VAE, scalable pre‑training strategies, and large‑scale data pipelines. Training employs a combination of DP, FSDP, RingAttention, and Ulysses parallelism, along with Context Parallelism for long sequences, achieving near‑linear scaling on multi‑GPU setups.

Data preparation involved a four‑step cleaning pipeline to construct a high‑quality, diverse image‑video dataset. Memory optimization combines layer‑wise offloading, fine‑grained gradient checkpointing, and PyTorch’s memory manager to reduce fragmentation.

During inference, the model uses CP‑based distributed acceleration and model‑splitting techniques to lower latency on multi‑GPU systems. Training stability is enhanced by Alibaba Cloud’s intelligent scheduling, slow‑node detection, and auto‑recovery, yielding a 98.23% successful restart rate with an average restart time of 39 seconds.

Overall, the open‑source release of Wanxiang 2.1 marks a milestone for Alibaba Cloud, completing a full‑modal, full‑scale open‑source ecosystem that spans from large language models to multimodal visual generators.