MiniMax launched MiniMax Design on August 20, a multimodal creation platform built around its in-house video model H3. Users describe what they want in natural language, and the system breaks down the script, lays out shots, generates keyframes, produces video, adds music, captions and edits — carrying the project through to a finished film.
The product is assembled from several pieces: a workflow canvas, a 3D director's console, an asset hub, a Skills library, and a set of specialized agents. The canvas keeps the relationships between assets and shots visible on the surface — a single edit doesn't require starting over, and you can branch off new versions and keep going. The 3D director's console is more literal: you place characters in a 3D scene, adjust camera angles and poses, and what you build becomes the storyboard reference that's then handed to the model for generation. The Skills library holds ready-made style templates — paper collage, ink wash, anime explainer, and so on — or you can build a custom Skill through conversation. Assets are saved locally automatically, and can be stored to the asset hub with one click. ComfyUI workflows can be imported directly.
What sits between 15 seconds and 3 minutes
H3's single-generation cap is 15 seconds. Among the examples MiniMax released are a 42-second short story, a one-minute music video, and a 3-minute documentary. Roughly speaking, three minutes means at least a dozen shots stitched end to end, with each shot's character likeness, wardrobe, lighting and camera language all needing to match, arranged into a timeline in narrative order.
That's the actual problem Design is solving. Over the past year, progress in AI video has concentrated on single-shot quality — higher resolution, more convincing physics, hands that no longer grow six fingers. But turning a dozen-plus individual shots into a watchable film runs on a different set of capabilities: character consistency, unified style, timeline organization, and the ability to survive a fifth revision without falling apart. MiniMax's division of labor has H3 handle only shot generation, agents handle unifying style and organizing the timeline, and a harness that checks back in at key points to ask what the core direction is, drawing on a specialized knowledge base behind the scenes for multi-round quality checks.
This structure follows almost the same template as today's AI coding tools: the underlying model handles single-point output, with an orchestration layer on top that can plan, self-check, and be interrupted by a human. Video has more variables — get one shot wrong and everything downstream has to change — so the verification step is pushed even further forward.
Whose business is this squeezing
MiniMax is unambiguous about the target scenarios: full-pipeline short dramas, bulk e-commerce image and video production, brand TVCs, promotional materials, explainer videos, and PVs/MVs. What these categories share is high volume, tight turnaround and low per-piece budgets — exactly the tier most easily overtaken by a more efficient tool chain.
According to iFanr, H3 plus Design has already cut into the professional creative workflow that Adobe has long held. That's half true. A professional editor's core asset was never just software features — it's judgment about material and the ability to communicate with clients, and agents can't take that over anytime soon. But once the repetitive grunt work at the base of a project — finding references, laying out storyboards, batch-generating images, syncing tracks, adding captions — gets packaged into a workflow triggered by a single sentence, the number of projects one person can run at once changes. What the industry calls "AI cutting content costs to a tenth" is mostly cutting this part.
MiniMax's string of moves
Tracing the timeline back makes this clearer. The H3 video model was open-sourced first, spreading generation capability widely; Hailuo AI built up scale in consumer-facing video generation domestically; now Design connects the two ends into a project-based workbench. Open source trades for ecosystem, the consumer end trades for data, the professional end trades for revenue — the three lines move in step.
For domestic content teams, the most immediate near-term change is probably the ComfyUI workflow import channel. This group has long since built up a pile of node graphs with parameters already tuned, and being able to bring them over as-is means not having to relearn everything from scratch. The biggest friction in switching tools rarely comes from not being able to learn the new one — it usually comes from not wanting to give up the thing that already works.
Sources: MiniMax official release, iFanr, CocoLoop editorial compilation; figures on H3's single-generation length cap and example film lengths are based on public demonstrations.