MiniMaxH3

The open weights are free — released under the MiniMax H3 Community License, royalty-free, for both research and commercial use (see Q3 for terms and Q4 for regional scope). Running locally costs you only your own hardware and electricity. The hosted API is pay-per-use (pricing), and you can also try H3 in the Hailuo AI app.

H3 generates video with synchronized stereo audio in the same pass — up to 15 seconds, up to 2K via the API (the open weights generate natively at 768p short-side). Two model variants cover every workflow: FL2VA (text-to-video, plus optional first and/or last frame control) and Ref2VA (omni-reference: up to 9 images, 3 video clips, and 3 audio clips as references — for consistent characters, video editing, motion reference, and clip continuation).

Yes. The Community License grants royalty-free commercial use, with three conditions to know about: (1) if your commercial products or services exceed US$20 million in yearly revenue, you need separate written authorization — contact api@minimax.io with the subject "MiniMax H3 licensing – authorization request"; (2) commercial products using H3 must prominently display "MiniMax H3" in their user interface; (3) all use must follow the Acceptable Use Policy, and if you host H3 for others you must implement reasonable content safeguards. You own the outputs you create and the derivative models you train. One notable restriction: outputs and weights may not be used to train or improve other AI models (distillation into non-H3 models is not permitted).

The open-weight license currently excludes the EU, UK, US, and South Korea while regulations for generative video (likeness, copyright, content safety) are still settling in those jurisdictions. This means "not yet," not "not ever." Two options today: the API is available globally with built-in safeguards, and organizations in excluded regions can apply for a formal license for reviewed deployments. We will expand the territory scope as the regulatory picture clarifies, and will announce any changes explicitly. (Full reasoning: License Q&A.)

The fastest path for individuals is ComfyUI, which has native day-0 support with bundled workflow templates: install ComfyUI, download three components (a diffusion model, the text encoder, and two VAEs) from the ComfyUI repackage into ComfyUI/models/, and load a MiniMax H3 template — step-by-step tutorial. For Python pipelines use the Diffusers release or DiffSynth-Studio; for low-VRAM machines see WanGP; for servers see Q7. The official GitHub repository includes deployment validation methods and reproducible request scripts to confirm your local setup matches reference behavior.

Less than you’d think. Community-verified reference points: 12 GB VRAM (RTX 3060-class) runs 480p clips with audio using the pruned int8 checkpoint (~42 GB of downloads); 16–24 GB is comfortable, minutes per clip once optimized; RTX 5090 / DGX Spark reach roughly 4× further with NVIDIA’s Sol Engine stack; Apple Silicon works via MLX (128 GB unified memory recommended) or the NF4 route from 16 GB; community stacks go as low as 5–8 GB. Plan for 32–64 GB of system RAM — layers that don’t fit in VRAM are staged from RAM. If you have no GPU, Colab GPU runtimes work.

Yes — two first-class serving stacks: vLLM-Omni (OpenAI-compatible video endpoint; official recipe, e.g. an 8.7 s clip end-to-end in ~87 s on 4×B300) and SGLang Diffusion (H3 on its supported list; API/CLI/SDK; runs on NVIDIA, AMD Instinct via ROCm, Apple MPS, and more). Remember the hosted-service obligations from Q3 if you serve third parties.

Same model family, three capability tiers. Open weights: generate locally at 768p short-side, full control, your hardware, free. The platform/API adds two server-side stages: H3-Context-IR, which refines your text/image/video/audio instructions into the representation the model consumes best (better results from the same prompt), and H3-Regenerate-2K, which takes a 768p result plus its original context and regenerates it at native 2K with finer detail. They compose: many creators generate and iterate locally for free, then send the keeper through the API for the 2K master. The repo README ships paired scripts showing the same job on both paths.

Sign in at platform.minimax.io, create an API key, and call the video generation endpoints — full references: H3-Context-IR and H3-Regenerate-2K. The model repo includes ready-to-run curl scripts for every task (T2VA / I2VA / Ref2VA, 768p and 2K paths) under scripts/readme/; 2K creation uses the video-generation-v2-create endpoint. Pricing: minimax.io/price.

Yes — the license explicitly covers derivative models, which you own. The community toolchain is already active: AI-Toolkit supports H3 (including a distillation training adapter), and the first character/style LoRAs, acceleration LoRAs, and full fine-tunes are on Hugging Face. Note the license requires modified-file notices when you redistribute, and derivatives remain under the Community License terms.

Prompt structure matters more on H3 than on most models — it reads your prompt with a full multimodal language model. Start with the two official guides: base (FL2VA) and reference (Ref2VA) — reference terminology from the guide measurably improves multi-reference results. If you work with Claude, Codex, or other coding agents, install the official prompt-writing skill from our GitHub repository (npx skills add https://github.com/MiniMax-AI/MiniMax-H3 --skill h3-prompt-writing — one of nine bundled skills, including eight style-specific ones). On the API, H3-Context-IR does this refinement for you.

The Acceptable Use Policy (Exhibit A of the license) applies to the weights, derivatives, and outputs: no unlawful use, no infringement of others’ IP or likeness rights, no harm to minors, no use of outputs to harm people, and — for anyone hosting H3 as a service — reasonable safeguards are required. The model’s capability is open; responsibility for use travels with it.

Join the official Discord community · code, skills, and deployment scripts: github.com/MiniMax-AI/MiniMax-H3 · weights and docs: Hugging Face / ModelScope · updates: @MiniMax_AI / @Hailuo_AI · API: platform.minimax.io · direct contact: model@minimax.io.