A14

Intel
2026-07-10 09:39:13

Intel Releases INT4 Quantized Versions of Alibaba's Wan2.2 Video Models, Drastically Reducing VRAM Needs

Intel has released three INT4 quantized versions of Alibaba's Wan2.2 video models on Hugging Face: T2V-A14B, I2V-A14B, and TI2V-5B. Using the AutoRound tool, weight precision dropped from 2 bytes (BF16) to 0.5 bytes, cutting VRAM requirements to about a quarter. The A14B models originally needed 80GB VRAM for 720p; after quantization, they could potentially fit on RTX 4090. Intel has not yet provided benchmark data on quality or memory usage. Users are advised to deploy using Intel's vllm-omni fork.

310
Intel Releases INT4 Quantized Versions of Alibaba's Wan2.2 Video Models, Drastically Reducing VRAM Needs