MiniMax H3 video generation that fits in 8 GB of VRAM

What it does

FreeVideo is a local inference engine for MiniMax H3, built on OpenVDN’s 8-step VDN-H3 model with Video DeltaNet’s hybrid attention. It runs as a ComfyUI plugin, with a Windows launcher that sets up the whole stack and command-line support on Linux. Inputs go beyond text: first/last frames plus image, video and audio references, and community LoRAs are supported.

The interesting bit

The memory story is the real trick. FreeVideo coordinates VRAM, system memory and disk together — weight streaming, async prefetching and chunked computation keep peak memory down to 8 GB VRAM / 16 GB RAM. It also probes your GPU and picks between native FP8 compute or FP8 storage with BF16 compute depending on architecture, and auto-selects available attention kernels.

Key highlights

• Hardware-adaptive execution: FP8 path chosen per GPU architecture, attention kernels probed automatically
• Low-memory inference via weight streaming, async prefetch and chunked computation
• ComfyUI workspace with two-pass sampling, batch generation, history, and a node view for finer control
• Windows launcher reuses existing ComfyUI installs and model folders; offline installation supported
• Apache 2.0 code; model weights under the MiniMax H3 Community License

Caveats

• The model weights carry territorial and acceptable-use restrictions — check the license before shipping anything commercial
• The Windows launcher is labeled a preview release
• Linux users get the command line only; the polished workspace experience is ComfyUI-centric

Verdict

Worth a look if you’ve got a modest GPU and want to run a serious video model locally without renting an A100. If you need permissive weights or non-ComfyUI workflows, read the license first and look elsewhere.