The first time an open-source video model goes head-to-head with closed-source flagships: MiniMax H3 is absolutely crushing it. In image generation, there's SD as open source; in large language models, there's LLaMA and DeepSeek open source—two major "open source kills closed source" battles that have driven API prices down to the floor. But the video generation space has been missing that moment. Previous open-source models were too far behind Kling and Seedance—they weren't even in the same league.

💡 What You Will Learn

The first time an open-source video model goes head-to-head with closed-source flagships: MiniMax H3 is absolutely crushing it. In image generation, there's SD as open source; in large language models

Open-Source Video Models Finally Go Head-to-Head with Closed-Source Flagships: MiniMax H3 Is a Game-Changer

Image generation has SD as an open-source option, and large language models have LLaMA and DeepSeek—two "open-source crushes closed-source" battles that drove API prices to rock bottom. But the video generation space has been missing that moment; previous open-source models lagged far behind Kling and Seedance, not even in the same league.

That changed today with MiniMax H3.

On August 3rd, MiniMax released H3 weights. With 33.1B parameters, it ranks #1 globally in video editing on Artificial Analysis (1130 Elo) and #2 in text-to-video. ComfyUI rolled out native support the same day with v0.30.0, including official workflows—no waiting in line for beta access, just update ComfyUI and run it locally.

H3 supports text-to-video, image-to-video, first-and-last-frame control, V2V motion transfer, and in-place editing. It also natively outputs stereo audio—the sound isn't post-produced; it's generated alongside the video by the model.

The API pricing is a knockout: 2K resolution at ¥0.8/second, roughly $7.80/minute. Seedance 2.0 costs $22.45, and Kling 3.0 is $20.16—H3 comes in at under a third of its competitors.

But what excites me most is local deployment. The original full-precision inference requires 123.6GB of VRAM, but ComfyUI prunes and int8-quantizes the modulated weights, cutting memory usage to 42.5GB—a 66% reduction. Combined with dynamic VRAM offloading, even an RTX 3060 can handle it. Reddit users have already tested it: a 5070 Ti 16GB with 32GB RAM generates a 720p 20-second video in just 4 minutes.

Honestly, I wasn't optimistic about "open-source video generation" before—video models are massive, with parameters in the tens of billions, and training and inference costs are dozens of times higher than image models. But ComfyUI actually made it run on a 3060. Any creator with a GPU can jump in now, and the ecosystem flywheel—LoRA communities, workflow marketplaces, model fine-tuning—is something closed-source models can never replicate.

From SD to LLaMA to H3, open source is challenging closed source for the third time, and this time it's video's turn. The last "from unplayable to playable" shift happened in image generation, and we all saw the results. What about this time?