MiniMax H3 dropped today, and it’s the company’s first video model with open weights. That’s a bigger deal than it sounds.
I’ve been watching the AI video space turn into a horse race over the last six months. ByteDance’s Seedance 2.5 landed last week with 30-second clips. Google’s Veo 3 keeps getting better. Runway, Pika, Kling, everybody’s shipping. But nearly all of them are closed. You rent access. You don’t run the model.
MiniMax H3 changes that math. The weights are on Hugging Face right now, under a permissive license. And ComfyUI shipped day-zero support with optimized inference that fits the model on a consumer GPU.
What MiniMax H3 actually does
H3 is an omni-modal video model. You feed it text, images, video, or audio and it generates video with real stereo sound in the same pass, not audio bolted on afterward. Output runs up to 2K resolution and 15 seconds per clip.
The input modes cover the spread most creators actually need:
- Text-to-video. Prompt only. The base mode.
- Image-to-video. Bring a still image to life. Good for animating concept art or product shots.
- First-and-last-frame. Pin the opening frame, the closing frame, or both and let the model fill in the middle. This is the filmmaking mode.
- Reference-to-video. Supply reference images, video, or audio and carry a subject, motion, or voice through the clip. Motion transfer (where a reference video provides camera movement while the subject and style come from elsewhere) is the one that matters most for workflow iteration.
The cross-modal understanding is what MiniMax leads with. Real projects rarely pull from one modality. You have a reference image, a style board, maybe a scratch audio track. H3 takes all of them together and resolves the relationships in one go.
Then there’s the audio. Every output comes with native stereo audio generated alongside the video. No separate TTS step, no lip-sync plugin, no post-processing. For a model at this price point (free, open weights), that’s unusual.
The ComfyUI optimization is the real story
Open weights are great. Open weights that run on consumer hardware are what actually matters.
The ComfyUI team did significant engineering to get H3 running on an RTX 3060. The model’s modulation weights (about 40% of the total parameters) were pruned and replaced with a functionally equivalent lookup table. Combined with int8 convolution quantization and custom kernels, the total memory footprint dropped 66%: from 123.6 GB at full precision to 42.5 GB.
With ComfyUI’s dynamic VRAM offloading, that fits on a 12GB consumer card.
I’ve watched enough model launches to know the pattern. Company drops weights, model requires 8x A100s to run, enthusiasts wait six months for a community quant that barely works. That’s not what happened here. ComfyUI had the optimizations ready at launch.
How it stacks up
H3 is MiniMax’s third-generation video model, following Hailuo 01 and Hailuo 02. It’s the first with open weights, and the first with multimodal context understanding that collapses five separate tasks (text, image, video, audio input, and video-plus-audio output) into one model.
The 15-second cap is shorter than Seedance 2.5’s 30 seconds, but H3 gives you permissive weights and local inference, which Seedance doesn’t. Different tradeoffs for different people. If you’re a studio with a render farm, Seedance’s longer clips might matter more. If you’re a solo creator who wants to iterate fast on your own hardware, H3’s local inference is the whole game.
The example outputs from the ComfyUI post are genuinely impressive. A comic-book-style short with synchronized dialogue and graphic overlays. A high-fashion editorial sequence with kintsugi masks, liquid ribbons, and a golden dragon. An editorial tech product film with duotone lighting and macro shots of a transparent gaming mouse. These aren’t generic “bear playing guitar” demos. They’re directed, composed, and styled.
What I’d keep an eye on
The open weights are real and the Hugging Face repo is live. That’s the headline. But a few things are still unclear:
The license terms matter. The ComfyUI post says “permissive” but I haven’t seen the exact license text yet. If it’s Apache 2.0 or MIT, commercial use is on the table. If it’s a custom research license, it isn’t.
The 15-second cap is real. Seedance does 30. For longer-form work, you’d need to chain clips or use a different tool.
And as always with day-zero launches, expect rough edges. The ComfyUI workflows are available now, but real-world testing on varied hardware will surface things the optimized benchmarks don’t show.
Still: open weights, native stereo audio, 2K output, runs on a 3060, day-zero ComfyUI support. That’s a strong launch. MiniMax just gave the open-source video community something it didn’t have last week.
