MiniMax H3 Drains the Last Moat in Open Video

For two years, the story of open-weight video generation has had a consistent shape: the open models — Wan, HunyuanVideo, LTX-Video, Mochi — kept closing the visual gap with the frontier, while audio stayed a closed-model moat. Veo 3 could give you a clip where the dialogue, footsteps, and room tone came out of the same forward pass. Open models gave you silence, and you bolted sound on afterward with MMAudio or a Foley model and hoped the sync held.

That's the moat MiniMax just drained, mostly. H3 — the model behind Hailuo 3.0, announced July 31 — generates up to 15 seconds of video at 2K resolution with native stereo audio in a single pass, and the weights landed on Hugging Face days later. ComfyUI shipped day-0 support in v0.30.0, with quantized variants that pull the footprint down far enough to matter. It's the first time anything close to Veo-3-class joint audio-video generation has been runnable on hardware you own.

It's also a release with an asterisk large enough that "open weights" deserves scare quotes for a lot of readers of this article. More on that below.

What actually shipped

H3 is a 33B-parameter dense, single-stream transformer — MiniMax's "H3-Omni Transformer" — paired with separate video and audio VAEs and, notably, Qwen3VL-32B as the text encoder. So the full pipeline is roughly 65B parameters before you generate a frame. About 13B of the transformer's parameters sit in AdaLN modulation branches, a detail that matters in a minute.

The weights come as two CFG-distilled, task-specific checkpoints: FL2VA handles text-to-video and first/last-frame conditioning, Ref2VA handles reference-driven generation. The reference story is the underrated part — H3 accepts images, video clips, and audio clips as references in one request, which is how you get consistent characters, motion transfer, and a voice that persists across shots. That's a capability class open models simply haven't had; it's closer to an editing system than a text-to-video toy.

To be precise about the "first" claim: open-weight joint audio-video generation technically existed — Character.AI's Ovi did 5-second 720p clips with sound last year — but nothing at this scale, resolution, or reference flexibility. Third-party gateways price H3 around $0.13 per second at 2K, roughly a third of comparable frontier models, and early leaderboard placements from Artificial Analysis put it near the top in video editing and text-to-video. Treat those rankings as provisional; the model's been public for three days.

Running it without an H100

In bf16, H3 is 123.6 GB — a non-starter for local use. ComfyUI's repackaged weights exploit that AdaLN detail: because modulation branches hold ~40% of the parameters, pruning them plus int8 quantization cuts memory about 66%, to 42.5 GB for the smallest variants. ComfyUI claims that, with its weight streaming, this runs on cards down to an RTX 3060. Believe the arithmetic, not the implication: streaming 40+ GB of weights through a 12 GB card will work and will be painfully slow. The realistic local floor is a 24 GB card with a lot of system RAM, and a comfortable experience starts at a 5090 or a rented A100/H100.

Adoption is genuinely turnkey by local-video standards: update to ComfyUI 0.30.0+, grab one of the three bundled workflow templates (text-to-video, image-to-video, reference-to-video), drop the diffusion model, Qwen3VL text encoder, and the two VAEs into their model directories, and queue. If you've run Wan 2.2 in ComfyUI, nothing here will surprise you except the audio outputs. For everyone else, the per-second gateway pricing is the sane on-ramp — at $0.13/second, a 15-second 2K clip costs about two dollars, which undercuts renting an H100 unless you're batching heavily.

Read the license before you ship

Here's the asterisk. The MiniMax H3 Community License allows free commercial use below $20M in annual revenue, with attribution — familiar Llama-style terms. But MiniMax's own license FAQ says self-serve open-weight use is temporarily not granted in the EU, UK, South Korea, and the US, citing regulatory uncertainty around video models — likeness rights, copyright, content safety. Organizations in those regions are told to apply for a formal license and demonstrate compliance controls.

Sit with that: the open-weight release of the year, for video, excludes the four jurisdictions where most of its would-be commercial adopters live. Enforcement against a local ComfyUI install is obviously theoretical, and research and personal use aren't the concern. But if you're building a product in San Francisco or Berlin, "download it from Hugging Face" is not currently a compliant path — the API and licensed gateways are. And the regulatory nervousness isn't abstract: MiniMax is still defending the lawsuit Disney, Universal, and Warner Bros. Discovery filed over Hailuo's rendering of copyrighted characters. A model that now generates those characters speaking raises the stakes of that case, not lowers them.

Where this lands

Technically, this is a genuine shift, not hype. Native audio was the last capability that cleanly separated closed frontier video from open weights, and the two-stage generate-then-Foley pipeline that every open workflow relied on is now legacy — MMAudio-style post-hoc sound survives only where you need audio for footage you didn't generate. Expect Wan and Hunyuan to answer with native audio within months; a moat this visible doesn't stay unique.

Practically, adopt in tiers. Hobbyists and researchers anywhere: this is your new default, and ComfyUI is the fastest way in. Commercial builders outside the carve-out regions: self-hosting at $20M-and-under terms is a real cost lever against per-second APIs. US/EU/UK/Korean companies: evaluate locally, ship through the API or a gateway, and watch whether MiniMax relaxes the territory clause once the regulatory picture — and the courtroom one — clears. The weights being on your disk has never been the same thing as the weights being yours to use.