Skip to content

← The Lab

FreeVideo is trending: local MiniMax H3 video on an 8 GB GPU, now with 48 and 60 fps export

FlashML's FreeVideo runs the MiniMax H3 video model on the computer you already own, starting at 8 GB of VRAM and 16 GB of RAM. It hit GitHub Trending and a big curator account in the same 24 hours, and v0.6.5 (Nov 19) adds 48 and 60 fps export with in-between frames made on your GPU.

What it is. FreeVideo is an Apache-2.0 launcher and ComfyUI workspace for MiniMax H3, built on FlashML's Video DeltaNet (VDN) work. Per the README, it gets peak memory down to as little as 8 GB of VRAM and 16 GB of system RAM by streaming weights, prefetching in the background and computing in chunks, and it picks precision and attention kernels for your card. Inputs include text, first and last frames, and image, video and audio references. Since October it has also had community LoRAs, four quality levels, minute-plus videos, a timeline editing view and agent control, so Claude Code or Codex can plan shots, submit jobs and hand back an MP4. v0.5.0 (Nov 10) added local still images up to 16 MP.

What v0.6.5 changes (Nov 19, 1:54 PM MST). The Export window now offers 48 and 60 fps. FreeVideo makes the in-between frames on the graphics card, leaves cuts and still frames alone, and copies the audio unchanged. Other changes: every generated or imported picture is also saved as its own file in the output folder (you can switch this off), there's a searchable LoRA menu grouped by folder, face close-ups are redrawn when a costume change leaves the old clothes, and failures now name the cause and the next step instead of a raw error code. Model folders can also live on another drive through symbolic links. Downloads are on the release page for Windows 10/11 with an NVIDIA GPU, Linux x86_64 with an NVIDIA GPU, and macOS 14+ on Apple silicon (the Mac build is still labeled a preview).

Why video teams should care. It's the cheapest way we've seen to put a current, audio-capable video model on a workstation you already own. That's handy for drafts, look tests and B-roll ideas you don't want to burn cloud credits on. The 8 GB number is the floor, not the comfortable spot: expect longer waits on small cards, and use the higher quality levels on bigger ones. One community user reported a 71-minute first setup and about CNY 6.69 of rented RTX 5090 time for four clips, which gives a rough sense of the first-day cost.

Read the license. The FreeVideo code is Apache-2.0, but the H3 weights it downloads are under the MiniMax H3 Community License, which has territorial and acceptable-use restrictions. Check it before you put H3 output in paid client work.

Signals: on GitHub Trending (Python, #16) on Nov 19 and again on Nov 20, with about 3,700 stars total. Tracked curator @RoundtableSpace posted it Nov 20 at 4:45 AM MST (about 148 likes and 48k impressions in the first hours). Trending plus a curator within 72 hours makes it HOT by our rule.

Related tools on ReelStack

Skills & repos in our directory

More from the Lab

Topics: ComfyUI · Open-weight video & audio models