MiniMax Design: The Video Model Finally Got a Director

MiniMax launched Design on August 20, and it is the most honest attempt yet at fixing what everyone pretends is fine about AI video. The underlying model, H3, generates 15 seconds per clip. That has not changed. What changed is what sits on top: a workbench where agents break down your script, lay out the shots, generate keyframes, produce the video, add music, captions, and edits, and hand you a finished film. MiniMax's own examples are a 42-second short story, a one-minute music video, and a three-minute documentary. Three minutes is roughly a dozen shots stitched together, with character likeness, wardrobe, lighting, and camera language all matching across them. That is the problem being solved.

The actual problem: continuity, not generation

For a year, AI video progress has been about single-shot quality: higher resolution, better physics, hands that stopped growing extra fingers. None of that helps you ship a film. A film is a dozen shots that agree with each other, arranged in narrative order, and still coherent after the client's fifth revision. MiniMax's division of labor is explicit: H3 handles shot generation only, agents handle style consistency and timeline organization, and a harness checks back at key points to ask what the core direction is.

If you have watched the coding-agent wave, this structure is familiar. The model produces single-point output, an orchestration layer plans and self-checks, and a human can interrupt at the keyframes. Video has more variables than code: get one shot wrong and everything downstream changes. So MiniMax pushed the verification step even further forward. The 3D director's console is the physical version of that: you place characters in a scene, set camera angles and poses, and what you build becomes the storyboard reference handed to the model. You are not prompting a video, you are blocking a scene.

What Design actually contains

The product is assembled from five pieces, per MiniMax's launch material:

• A workflow canvas that keeps the relationships between assets and shots visible, so one edit does not force a restart, and you can branch off new versions.
• The 3D director's console described above.
• An asset hub with one-click saving, plus automatic local saving.
• A Skills library: ready-made style templates (paper collage, ink wash, anime explainer) or custom Skills built through conversation.
• Specialized agents for copy, image, video, and audio, each handling its lane.

One detail matters for people already working in video: ComfyUI workflows can be imported directly. If you have a pile of tuned node graphs, you bring them over as-is instead of rebuilding them. The biggest friction in switching tools is usually not learning the new one, it is giving up the thing that already works.

Who gets squeezed

MiniMax names the targets directly: full-pipeline short dramas, bulk e-commerce image and video production, brand commercials, promotional material, explainers, music videos. These categories share high volume, tight turnaround, and low per-piece budgets, exactly the tier a more efficient tool chain overtakes first.

Chinese tech media have already written the "Adobe is threatened" version of this story. That is half right. A professional editor's asset was never software features alone, it is judgment about material and the ability to talk to clients. Agents are not taking that over soon. But the grunt work at the bottom of a project, finding references, laying out storyboards, batch-generating images, syncing tracks, adding captions, can now be packed into a workflow triggered by one sentence. The number of projects one person can run at once changes. What the industry calls cutting content costs to a tenth is mostly cutting this part.

There is a strategy under this release too. H3 was open-sourced first, which spreads generation capability everywhere. Hailuo AI built the consumer scale. Now Design connects the two ends into a project-based workbench: open source buys the ecosystem, the consumer app buys the data, the professional tool buys the revenue. Three lines, one trajectory.

What we could not verify

Two things are not in this review. Pricing: MiniMax had not published a clear price sheet for Design at the time of writing, so this is not a cost review yet. And hands-on testing: we have not run a real project through it. The example films are MiniMax's own demonstrations, which is exactly the claim you should treat with skepticism. Our verification policy means we will update this page when we have a production bill and a finished, non-MiniMax project. Until then, treat Design as the most promising video pipeline architecture of 2026, and still unproven at scale.

If you want to compare the underlying models instead of the pipeline, our video generators guide covers the alternatives, and the Seedance card is the current value pick. Facts in this review were checked against MiniMax's launch material and independent reporting from Cocoloop and iFanr. Last verified August 22, 2026.