Find your next open-source project

Search by name, purpose, or keyword — or paste a GitHub repository URL

FlashML-org/FreeVideo

Machine Learning

Runs MiniMax H3 video generation locally on consumer GPUs with as little as 8 GB of VRAM and 16 GB of RAM.

3,379
448
Python
Apache-2.0
2 months old

What it does

FreeVideo is a local inference engine for the MiniMax H3 video model, built on OpenVDN's 8-step VDN-H3 model and Video DeltaNet's hybrid attention. It targets consumer hardware: the project claims generation can run with 8 GB of VRAM and 16 GB of RAM by coordinating VRAM, system memory and disk.

The engine picks weight format, compute precision and attention kernels based on the GPU — int8 on GeForce and RTX 30-series cards, FP8 on workstation and datacenter cards — and uses weight streaming, asynchronous prefetching and chunked computation to keep peak memory down. Inputs include text prompts, first and last frames, and image, video or audio references. It runs as a ComfyUI plugin with a dedicated workspace, plus node view for LoRAs and custom workflows.

Getting started

The Windows path is the simplest: download the FreeVideo.exe launcher, choose an existing ComfyUI folder or install a new one, point it at any existing model folders to reuse, then click Install & launch. ComfyUI opens in the browser with the FreeVideo workspace; missing models are downloaded automatically.

On Apple silicon there is a preview via the macOS DMG, tested on an M5 Mac with 24 GB of unified memory. It is not notarized yet, so macOS blocks the first launch; the first-open steps explain how to approve only this app. Linux users can run the AppImage or set up from a terminal or server with git clone and ./setup.sh; the Linux guide covers server and SSH use.

If ComfyUI is already installed, add FreeVideo as a custom node:

cd ComfyUI/custom_nodes
git clone https://github.com/FlashML-org/FreeVideo.git

Restart ComfyUI, open Workflow → Browse Templates → FreeVideo → FreeVideo-All-in-One, and complete setup in FreeVideo Settings. Update with git pull in custom_nodes/FreeVideo and restart.

Requirements and limits

An existing ComfyUI installation is not required — the launchers install ComfyUI, the runtime and models — but a fresh install downloads several model packs: the video model, text encoder, video and audio decoder, and sampling caches. For offline installation, the README links the packages from Quark and has you drag the ZIP files into the launcher without extracting. A server install needs network access or SSH setup, and LAN access uses freevideo server --listen 0.0.0.0.

The code is Apache-2.0, but the model weights carry the MiniMax H3 Community License with territorial and acceptable-use restrictions, so check those terms before commercial use. The Mac preview has a narrower hardware base than the GPU paths and is explicitly labeled preview.

The Windows launcher is a precompiled executable; the README says it is code-signed and suggests verifying its SHA-256 against the release page if security software blocks it. That verification step is the recommended way to confirm you have the intended binary rather than a modified copy.

HotRepo's take

On the 2026-11-19 Python daily list, FreeVideo recorded 3,352 stars and 445 forks with 99 stars gained and a best position of 15. That is a single-snapshot observation on the language-specific daily list, so it says who noticed the project that day, not whether its low-VRAM claims hold on a given GPU. The useful signal for prospective users is the mix of qualities the README documents: adaptive weight formats and kernel probing are described concretely enough to inspect in the execution-planning docs, and community-generated clips are collected in the results page. Verifying those against your own hardware is the practical next step.

Written by AI from the project's README, updated . The project's own docs are the source of truth.

You might also like

  1. Video-Rebirth/hyperflow-h3-flashA 13.7B few-step MiniMax-H3 derivative that generates video with synchronized stereo audio, in INT8 and GGUF formats.Python111
  2. MiniMax-AI/MiniMax-H3-ImageOpen-weights 16-megapixel image generation and editing model with a 33.1B DiT and 6-step sampling.117
  3. Betelgeuse-AI/BetelgeuseA C++/CUDA llama.cpp fork that streams MoE experts from NVMe to run large models on a 12 GB RTX 3060.C++111
  4. huangserva/h3-fightA Chinese prompt library, validator and ComfyUI runner for MiniMax H3 fight-scene video generation.Python104
  5. wici-ai/inferssdRun large mixture-of-experts GGUF models from SSD when they exceed your RAM or VRAM.C++160