Intel has announced the release of three INT4 quantized versions of Alibaba's Wan2.2 video generation models on Hugging Face. According to Haihao Shen, Intel's Chief AI Engineer, the models include T2V-A14B (text-to-video), I2V-A14B (image-to-video), and TI2V-5B (text-image-to-video hybrid), all quantized using the AutoRound tool.
Quantization Details
The process compresses each weight from 2 bytes in BF16 format to 0.5 bytes in INT4, reducing the weight size to approximately one-quarter of the original. The A14B models originally featured a Mixture-of-Experts (MoE) architecture with 27 billion total parameters and 14 billion activated per step, requiring at least 80GB of VRAM per GPU for 720p resolution. With INT4 quantization, VRAM demand could theoretically drop to around 20GB, significantly lowering hardware barriers.
Model Architecture and Performance
The TI2V-5B dense model could already run 720p at 24fps on a single RTX 4090 GPU in its original form. Post-quantization, its memory footprint further decreases and may support faster inference. However, Intel has not yet provided benchmark data on VRAM usage or visual quality, making third-party verification essential.
Deployment and Compatibility
Intel directs users to its proprietary vllm-omni branch for deployment, as these models do not work with the main vLLM inference pipeline. This requirement means developers must switch to Intel's maintain fork to leverage the optimization, potentially limiting broader adoption but underscoring Intel's strategic push into AI inference ecosystems.
Industry Implications
Video generation models are gaining traction, yet their high VRAM demands hinder deployment on consumer-grade hardware. Intel's INT4 quantization offers a viable path to reduce costs. If independent tests confirm acceptable quality loss, these quantized models could bring video generation to edge devices and SMBs. Investors should note the lack of independent evaluations and the need for a dedicated inference framework, leaving actual performance unconfirmed.