FlashML-org released FreeVideo, a local inference engine that runs the MiniMax H3 video generation model on consumer GPUs with as little as 8 GB VRAM and 16 GB RAM. Built on OpenVDN's 8-step VDN-H3 model using Video DeltaNet's hybrid attention mechanism, FreeVideo implements hardware-adaptive execution: it selects int8 quantization for GeForce and RTX 30-series cards and FP8 for workstation and datacenter GPUs, while automatically probing available attention kernels. Low-memory inference is achieved through weight streaming, asynchronous prefetching, and chunked computation. The system supports multimodal inputs including text prompts, first and last frames, and image, video, and audio references, plus community LoRAs. FreeVideo operates as a ComfyUI plugin with a dedicated workspace supporting two-pass sampling, batch generation, and history tracking, with a node view for finer control. One-click launchers for Windows, macOS (Apple Silicon, tested on M5 with 24 GB unified memory), and Linux handle ComfyUI setup, runtime environment, and model downloads, with offline installation via Quark packages. On an RTX 3090, int8 acceleration yields approximately 2.3x speedup for 5-second video generation. The code is Apache 2.0 licensed; model weights fall under the MiniMax H3 Community License with territorial and acceptable-use restrictions. Citation requested for the Video DeltaNet paper (arXiv:2609.20744).

Read original