MiniMax H3 Open Source: What Creators Should Know

Last updated: August 17, 2026. MiniMax H3 is now available as open weights — a general-purpose video model that accepts text, images, video, and audio in one context and outputs up to 15 seconds of 2K video with native stereo audio. This article explains what shipped, what did not, and how to start without wasting a weekend on the wrong checkpoint or license assumption.

---

TL;DR

---

What open-sourced on August 3, 2026

MiniMax released two task-specific checkpoints on Hugging Face, each with its own processor, tokenizer, text encoder, Visual VAE, and Audio VAE:

Both checkpoints generate native stereo audio in the same pass — not a separate audio model bolted on afterward. That is why H3 prompts need soundscape and music sections, not a single line saying “cinematic music.”

What is still not open source

Some modules remain API-only today:

If you self-host weights, plan for 768P-class local output unless you pipe results through the API regeneration endpoint. Read the official open-source announcement for the latest module status.

---

Why the prompting bar is higher

Unlike earlier text-to-video tools, H3 expects an audiovisual timeline, not a one-line scene description:

• Timed shot lists with camera verbs and cut points (At 00:03.000…)
• Separate soundscape and music (overall_soundscape, non_diegetic_music)
• Explicit reference roles in Ref2VA — every file needs a job (face lock, product color, camera move only)

Prompts under ~200 characters often fail on T2VA. Ref2VA needs the full six-section official structure. The limit is 7,000 characters — enough for a complete commercial spot with sound design.

That is why search interest spiked for “minimax h3 prompt” — creators need copy-ready templates and a compiler, not another generic “best prompts” list.

For the latest model landscape, see LTX 2.5 vs MiniMax H3 — speed, licensing, and when H3’s Ref2VA format still wins.

---

Three input modes to know

I2VA and L2VA are keyframe variants under the FL2VA checkpoint — align to Picture 1 at 0s, or converge to a supplied last frame.

Ref2VA capacity (API parity reference)

When writing Ref2VA prompts, official API limits are a useful ceiling even for local work:

Start with 2–4 references until results are stable. More files means more roles to declare — not automatic quality.

---

License — read before commercial local deploy

H3 ships under the MiniMax H3 Community License Agreement. MiniMax defines an Applicable Territory that excludes the United States, United Kingdom, European Union, and Republic of Korea for authorized use of the open weights (linked to ongoing generative-video litigation).

If you are a US/EU/UK/KR company planning to run weights locally, read the official license Q&A before you ship. Many creators in those regions still use H3 via API (MiniMax, Hailuo, fal.ai) under separate terms.

This is unrelated to MiniMax H3 Prompt Hub — we are an unofficial resource. Always verify MiniMax’s license page for current wording.

---

Hardware and runtime expectations

H3 is not a speed-first model locally. Community reports in August 2026 commonly cite:

ComfyUI workflows landed immediately after open source. Diffusers and SGLang paths exist for advanced users. If your goal is fast iteration on a consumer GPU, read our LTX 2.5 comparison — many creators run H3 for reference fidelity and LTX for speed.

---

How to start prompting (three paths)

Path A — Prompt Studio (fastest for format)

1. Open Prompt Studio
2. Pick Ref2VA or T2VA (or a Hub skill template)
3. Fill the brief — use free AI assist for signed-in users
4. Compile the official export and copy to Hailuo or the MiniMax API

Path B — Copy a commercial template

Browse the prompt library for product ads, brand spots, and explainers with teaching notes — not just a copy button.

Path C — Self-host weights

1. Download the checkpoint you need from Hugging Face (FL2VA/ or Ref2VA/)
2. Load the matching ComfyUI workflow
3. Paste a compiled export from Path A — do not paste a one-line ChatGPT blurb

Start with the prompt formula guide, then Ref2VA roles if you have product or talent assets.

---

What we ship on this site

Free to browse, compile, and use AI assist. Paid plans add credits for in-site video generation only.

---

FAQ

Is the prompt library free?

Yes — every guide and template on this site is free to read and copy. Compiling official exports and AI assist are free for signed-in users. Credits apply only to optional in-site video generation.

FL2VA or Ref2VA — which checkpoint do I download?

Ref2VA if you have reference images, video, or audio to lock. FL2VA for pure text, first-frame, or first+last-frame workflows.

Can I get 2K locally from open weights?

The open checkpoints target the base generation path. Regenerate-2K remains an API module — plan for 768P-class local output or use the API upscaling workflow documented by MiniMax.

How is this different from Hailuo?

Hailuo is MiniMax’s consumer product surface. This site focuses on the official export format (Ref2VA six sections, T2VA three fields) so your prompt survives API, Hailuo, ComfyUI, or self-hosted inference without rework.

Where are the official docs?

• MiniMax video prompt gallery
• Video generation API guide
• MiniMax-H3 on GitHub — includes the h3-prompt-writing skill

---

Next steps

1. Prompt formula guide — what to write, what to skip
2. Commercial template — copy and adapt
3. LTX 2.5 vs MiniMax H3 — if you are choosing a local stack

Unofficial resource — not affiliated with MiniMax.