MiniMax shipped H3 (also branded Hailuo 3.0) on July 31, 2026, and the open weights landed on August 5. It's an omni-modal video model: one transformer that takes text, images, existing video, and audio as inputs and returns a synchronized 2K video clip with native stereo audio, all in a single generation pass. That last part is what's actually interesting.
What Omni-Modal Means in Practice
Most video generation pipelines I've seen people build have at least two stages. You generate the video clip. Then you add audio separately, either with a different model, a music library, or a post-processing step. That's workable for silent B-roll or concept visualization. But once you want something closer to a finished video segment, audio-video sync becomes a problem you have to solve explicitly.
H3 sidesteps this by treating audio as a first-class output, not an afterthought. You can pass in a reference audio clip, a voice note, ambient sound, or nothing at all. The model generates video with synchronized stereo audio from the same attention pass. When people are speaking, mouth movements cohere with the audio. The soundscape matches the visual motion. No separate sync step. No stitching.
The input spec is flexible: up to 9 reference images, 3 reference video clips, and 3 reference audio clips per request. Clip length runs from 4 to 15 seconds natively, extendable to around 30 seconds with the Extend tool. Aspect ratios cover 21:9 to 9:16.
The Open Weights Fine Print
The weights dropped August 5 under the MiniMax H3 Community License. The terms: free for non-commercial use, and commercial use is permitted for organizations under $20M annual revenue with attribution required. Anyone above that threshold needs to negotiate a separate agreement.
If you're a startup or independent builder, this is basically open access. If you're enterprise, it isn't.
One real tradeoff for self-hosting: local inference caps at 768p. You also need H100-class GPU hardware to run H3 at all. The API delivers native 2560x1440 (2K). The resolution gap between 768p local and 2K cloud is visible, not subtle. So self-hosting gives you cost control and data privacy, but the API is the path to the best output quality. You can't fully have both right now.
The Pricing Math
MiniMax prices the API at $0.09/second at 768p and $0.13/second at native 2K, accessible through platform.minimax.io, OpenRouter, and EvoLink. A 15-second 2K clip comes out to about $1.95. Extend to 30 seconds and you're at roughly $3.90.
Comparable frontier video APIs were pricing 2.5 to 3x that per second of output before H3. For high-frequency generation scenarios, that compression matters.
The Artificial Analysis Video Arena leaderboard as of the launch date had H3 at #1 for text-to-video, #2 for video editing, and #3 for image-to-video with audio. Leaderboards in this space move fast and I'd weight them accordingly, but that's a strong opening position.
What This Changes for Builders
Two things I'd actually act on here.
First, if you're building a video generation workflow, dropping the separate audio stage simplifies your pipeline in a way that's easy to underestimate. Fewer models in a chain means fewer failure modes, fewer latency steps, and less state to manage between calls. A single call returning a coherent video-with-audio is architecturally cleaner than orchestrating two or three separate generation steps.
Second, the multi-reference input is interesting for agentic workflows. An agent collecting voice notes, example images, and short video clips from a user can bundle all of them as H3 references and get back a video that draws from the full set simultaneously. I haven't tested H3 at scale, and I don't yet know how it handles conflicting references or degrades when you push the reference count to its limits. That's worth benchmarking before committing to a production design.
At 4-5 second clips, you're looking at roughly $0.40-$0.65 per generation. That price point opens up video generation for use cases where per-call cost was the blocker before.
What to Keep an Eye On
The $20M revenue ceiling in the community license is the wildcard. MiniMax priced this to drive adoption, but you're building on a model where commercial rights depend on your revenue staying below a threshold. If you're building something that might scale past that point, factor in the renegotiation risk early.
The 768p local ceiling is also worth flagging. If your use case requires 2K output quality and on-prem inference for privacy or regulatory reasons, H3 doesn't fully serve both right now. Something to watch for in subsequent releases.
The MarkTechPost coverage from August 1 has the technical spec breakdown if you want to go deeper before integrating.