FlashML Runs MiniMax H3 Video AI on 8 GB GPUs

FlashML-org has released FreeVideo, an Apache 2.0-licensed local inference engine designed to run a derivative of the MiniMax H3 video generation model on consumer-grade hardware. By utilizing aggressive weight offloading, reduced precision, adaptive attention kernels, and streaming, the engine lowers the hardware barrier to just 8 GB of VRAM and 16 GB of system RAM. This is a significant shift from typical open video models, which usually demand a 24 GB GPU or reliance on hosted cloud APIs.

The engine is built on OpenVDN's eight-step VDN-H3 checkpoint, which employs Video DeltaNet hybrid attention. While the original MiniMax H3 is a 2K-capable multimodal model, the VDN-H3 checkpoint used by FreeVideo outputs video with a 768-pixel short edge, translating to roughly one megapixel depending on the aspect ratio. It generates clips lasting between 5 and 15 seconds at 24 frames per second. Because the system streams offloaded weights directly from disk to fit into smaller memory footprints, SSD performance becomes a critical factor, resulting in longer generation times and heavier disk traffic.

For AI practitioners and creators, FreeVideo integrates directly into existing workflows. It ships as a ComfyUI plugin, alongside a Windows one-click launcher and Linux command-line interface support. The tool supports community LoRAs, two-pass sampling, batch generation, and reusable adapter caches. Crucially, MiniMax H3 is a multimodal pipeline that accepts text, images, video, and audio to generate both video and native stereo audio simultaneously, meaning sound effects and music are baked into the output rather than added in post-production.

Currently, the project has garnered 546 GitHub stars and is being actively patched to resolve low-VRAM edge cases. By bringing local, multimodal video generation to mainstream consumer GPUs, FreeVideo democratizes access to advanced media synthesis tools that were previously restricted to enterprise-grade hardware.