About this project

OpenVideo is an open-source project that brings local video generation to your own hardware, positioning itself as "Ollama for video models." It provides a simple command-line interface (CLI) that follows the familiar `install · pull · run` workflow, allowing users to download and run the MiniMax H3 video model locally on NVIDIA GPUs. The project also includes a skill harness that enables coding agents (such as Claude Code, Cursor, Codex, and OpenCode) to generate high-quality videos by crafting official H3 prompts and validating them against hard constraints. Key features of v0.1.0 include: - **Local generation**: Run MiniMax H3 via ComfyUI on your own GPU, with support for different quantization tiers (INT8, W4, NF4) depending on VRAM. - **CLI**: Commands like `open-video pull h3`, `open-video status`, and `open-video run` provide an Ollama-like experience. - **Agent skill**: The `skill/h3-video/SKILL.md` file enables any agent host to generate videos end-to-end, with prompt grammar and validation built in. - **Judge loop**: An optional VLM-based judge can automatically refine low-scoring shots, with an honest `SKIPPED` fallback when no VLM is configured. - **Plugin architecture**: Designed to be extensible, with backends for new models, engines, and judges. OpenVideo is not a foundation model itself; it integrates with ComfyUI as the runtime engine. The code is licensed under Apache-2.0, but the MiniMax H3 model weights are covered by a separate community license with territorial and commercial-use restrictions. The project is currently in early development (v0.1.0) and focuses on single clips; multi-shot film planning is experimental.