The latest release of flashml-org/freevideo is v0.4.12, a Python inference engine that runs the MiniMax H3 video model locally to generate video with ComfyUI. The provided information lists no specific changes in v0.4.12. The project description states that it targets configurations down to 8GB VRAM and 16GB RAM.
The core aim of this project is running video models on consumer GPUs. It builds on OpenVDN's 8-step VDN-H3 and Video DeltaNet hybrid attention, automatically selecting int8 for GeForce and RTX 30 series and the FP8 weight format for workstation and data-center cards, and searching for attention kernels. It manages memory usage with weight streaming, asynchronous prefetch, and chunked computation, and takes text, first and last frames, and image, video, and audio references as inputs. The README reports that in v0.3.0, 5-second video generation on an RTX 3090 became about 2.3 times faster, and it explains that v0.4.0 added minute-long video, timeline editing, and agent calls such as Claude Code and Codex.
The performance evidence is limited to self-measurements on an RTX 3090 and a community results document; no independent verification is provided. There is also information indicating that there is no tests directory at the repository root and only test.sh can be found, so the scope of the test suite is unclear. Therefore, the memory handling approach targeting 8GB VRAM and 16GB RAM is worth referencing, but the evidence is insufficient to accept the performance figures as general results.