Intel has announced the release of three INT4 quantized versions of Alibaba's Wan2.2 video model family on Hugging Face, led by Intel AI engineer Haohao Shen. The move is aimed at significantly reducing the video memory required for inference, making large-scale video generation models more accessible on consumer hardware.
The released models include T2V-A14B (text-to-video), I2V-A14B (image-to-video), and TI2V-5B (text-image-to-video hybrid). All are quantized using Intel's AutoRound toolkit, compressing each weight from 2 bytes in BF16 format down to 0.5 bytes in INT4, slashing the total weight size to approximately one-quarter of the original.
Model Background and VRAM Requirements
The Wan2.2 series is Alibaba's major foray into video generation. The A14B variants use a Mixture-of-Experts (MoE) architecture with a total of 27 billion parameters and 14 billion activated per step. In their original form, running at 720p resolution required at least 80GB of GPU VRAM. After quantization, the theoretical VRAM demand could drop to about 20GB, potentially enabling operation on high-end consumer GPUs such as the RTX 4090.
The TI2V-5B, a dense model, could already run at 720p, 24 fps on an RTX 4090 in its original BF16 version. With INT4 quantization, VRAM usage will be further reduced, broadening accessibility.
Quantization Effectiveness and Deployment Guidance
Although Intel claims a dramatic reduction in memory footprint, the company has not yet released any benchmark data for VRAM usage or output image quality after quantization. Third-party verification is therefore essential for users to understand the real-world trade-offs.
For deployment, Intel recommends using its dedicated vllm-omni fork, as the quantized models are not integrated into the main vLLM inference pipeline. This fork is optimized for INT4 model loading and inference, ensuring better performance on Intel hardware.
This release marks a significant step by Intel in advancing AI model quantization and deployment, potentially lowering the barrier for running large-scale video AI models on consumer-grade equipment.