What MiniMax Shipped
MiniMax H3 — the third generation of the Hailuo video line — landed on Hugging Face as MiniMaxAI/MiniMax-H3 on August 3, 2026, under the MiniMax H3 Community License Agreement. The headline is not raw resolution or clip length. It is consolidation: one model that accepts text, images, video clips, and audio in any combination, and returns video with synchronized native stereo sound.
That is work that has until now been split across a stack of specialised models — one for text-to-video, another for image-to-video, another for video editing, another for a soundtrack, another to upscale. H3 folds those into a single set of weights you can download.
The Architecture, Briefly
H3 is built on what MiniMax calls the H3-Omni-Transformer: a 33B-parameter dense, single-stream Transformer using three-dimensional multimodal rotary position embeddings (MM-RoPE) to place text, image, video and audio tokens in a shared space. Roughly 13B of those parameters sit in AdaLN-related branches, and MiniMax notes those modulation outputs can be precomputed and cached — so an inference-only deployment does not need to load them.
Dense, not mixture-of-experts, is a deliberate choice for a video model: every token goes through the same weights, which keeps the audio and video streams tightly coupled instead of routed apart.
Two Checkpoints, Two Jobs
- H3-Base FL2VA — text plus optional first and/or last frame images. This is the classic text-to-video and keyframe-interpolation path.
- H3-Base Ref2VA — reference-driven generation. It accepts text plus up to 9 reference images, 3 reference video clips, or 3 reference audio clips, capped at 12 files total.
Ref2VA is the interesting one. Feeding a model an existing clip, a voice sample, and a few character stills in one request is what makes editing and continuity work possible, rather than rolling the dice on a fresh generation each time.
Output Specs
- Duration: 4–15 seconds at 24 FPS
- Resolution: shorter side defaults to 768px; 2K available through H3-Regenerate-2K
- Aspect ratios: 21:9, 16:9, 4:3, 1:1, 3:4, 9:16
- Audio: native 32 kHz stereo, with independent channel processing
- Languages: stable support for 11 — Arabic, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian, Spanish
Where It Ranks
On Artificial Analysis leaderboards, H3 takes first place in Video Editing with an Elo around 1130, edging out Gemini Omni Flash at roughly 1122, and places second in text-to-video and third in image-to-video. Video editing is exactly the category the Ref2VA checkpoint targets, so the ranking is consistent with what the architecture was built to do — and worth reading as a category win rather than a claim of overall video supremacy.
What You Can Actually Run Yourself
“Open weights” here means genuinely downloadable, but not casually runnable. MiniMax documents SGLang deployment across 4 GPUs, and community ComfyUI workflows and quantised GGUF conversions appeared quickly. A 33B dense video model with audio decoding is a workstation-or-cluster proposition, not a laptop one, and 2K output routes back through MiniMax's hosted API rather than the open checkpoints.
So there are two honest paths: rent it through the MiniMax Open Platform API if you just want clips, or self-host if you need the weights under your own control and have the hardware to match. The license is MiniMax's own community license, not Apache or MIT — read the terms before building a commercial product on top of it.
What This Means If You Run an Agent
An always-on agent does not need to host a video model to use one. The practical pattern is that your agent calls the MiniMax API when someone asks it for a clip, the same way it calls a search or image tool — see our guides for MiniMax on OpenClaw and MiniMax on Hermes Agent, and making videos with your bot for the end-to-end brief-to-MP4 flow. For H3 specifically — which endpoint ids to use, what a clip costs, and why a fal key rather than a MiniMax key is the working route today — see MiniMax H3 on OpenClaw.
What H3 changes is the shape of the request. A model that takes reference images, a reference clip, and a voice sample together means an agent can carry continuity across generations — same character, same voice, same look — instead of treating every clip as an isolated prompt. That is the difference between a novelty generator and something that can produce a consistent series.
The Limits Worth Knowing
Fifteen seconds is a hard ceiling, so anything longer is a stitching job you assemble yourself. 2K needs the hosted API. Availability is restricted in some regions. And as with every video model, the gap between a good demo prompt and a reliably repeatable result is still wide — budget for iterations.
Still, an open-weight model that unifies four input modalities and ships with native stereo audio is a real shift in what is downloadable rather than rented. If you want an agent that can act on that — on Telegram, Discord, WhatsApp, WeChat, or the web — OpenClaw Launch runs one with your own keys and your choice of model, no rebuild required.