FreeVideo lets home PCs with 8 GB of VRAM run MiniMax H3 video generation

FlashML has released FreeVideo, an open-source tool that runs MiniMax H3, a video generation model, on home hardware. The repository, which lives on GitHub, says it works on as little as 8 GB of VRAM and 16 GB of system RAM, adapting how it uses memory and compute to whatever the machine has. The project runs as a plugin inside ComfyUI, the node-based interface many local AI users already rely on, and accepts text prompts, first and last frames, and image, video, and audio references as inputs. The code is released under the Apache 2.0 license, while the MiniMax H3 model weights ship under a separate MiniMax community license, a split worth reading before building anything commercial on top of it.

The memory savings come with a clear trade-off. To fit a large model into 8 GB of VRAM, FreeVideo streams weights in and out of system memory and disk as it goes, which means heavy disk traffic and noticeably longer generation times. Community benchmarks put a ten-second clip at 1344 by 768 on an RTX 4060 Ti 16 GB at roughly nine minutes, and a contributor’s test on an 8 GB version of the same card took closer to ten. Patience is part of the deal: this is a tool for producing a few clips at a time, not for iterating through dozens of variations in an afternoon.

Version 0.2.0 landed on October 5 with four quality levels, Light, Medium, High, and Max, where each step up trades more generation time for better output. Earlier, version 0.1.2 added support for Apple silicon Macs, widening the hardware range beyond Nvidia cards. It fits a wider pattern of this year’s open models getting pushed onto ordinary consumer machines, the same direction video research has been moving on the training side.