FreeVideo is a local inference engine for running the MiniMax H3 video generation model on consumer GPUs. It is built on OpenVDN's 8-step VDN-H3 model, which uses the hybrid attention of Video DeltaNet (VDN); the VDN-H3 weights and MiniMax H3 base model come from those upstream projects, while FreeVideo supplies the runtime around them. The engine coordinates VRAM, system memory, and disk, adapting weight placement, compute precision, and attention kernels to the available hardware. For each GPU architecture it chooses either native FP8 compute or FP8 storage with BF16 compute, and it probes which attention kernels are available. Weight streaming, asynchronous prefetching, and chunked computation keep peak memory low, allowing inference with as little as 8 GB of VRAM and 16 GB of RAM. Inputs include text prompts, first and last frames, and image, video, and audio references, and community MiniMax H3 LoRAs can be used in a workflow. FreeVideo runs as a ComfyUI plugin that adds a dedicated creative workspace with two-pass sampling, batch generation, and a history of past creations, plus a node view for adding LoRAs or customizing the workflow. On Windows, a launcher (published as a windows-preview pre-release) sets up ComfyUI, the runtime environment, and the models, reuses existing model folders, and supports offline installation; the offline packages include model packs for 30/40-series or 50-series GPUs. Existing ComfyUI users can clone the repository into custom_nodes and open the FreeVideo-All-in-One template. On Linux, setup.sh installs it and a freevideo generate command renders a video from a prompt file to an MP4. The code is Apache-2.0 licensed, but the model weights fall under the MiniMax H3 Community License, which includes territorial and acceptable-use restrictions.