Open-Source Video AI Just Beat Sora

Open-source AI video generation just did something that was supposed to be impossible. It beat the closed models at their own game.

On August 3, Chinese AI startup MiniMax released the weights for its H3 video generation model. Hours later, the model sat at the top of the Artificial Analysis leaderboard -- the first open-weight model to reach that position in any major AI video ranking. It ranks first in video editing, second in text-to-video, and third in image-to-video, outperforming OpenAI's Sora, Google's Veo, and ByteDance's Seedance in multiple categories.

The Sora moat is no longer a moat. It is a speed bump.

What H3 actually does

H3 is not a video model in the conventional sense. MiniMax calls it an "omni-modal" model, meaning it processes text, images, video, and audio as a unified input space rather than running separate sub-models. The architecture is a 33-billion-parameter transformer that accepts up to nine reference images, three video clips, and three audio clips in a single prompt. It outputs four- to fifteen-second clips with native stereo sound generated alongside the video, not stitched on afterward.

The model card on HuggingFace reports 5,391 downloads within hours of release and 1,277 likes, making it one of the fastest-adopted video models on the platform. The community port for ComfyUI hit 399 likes, indicating strong early interest from the open-source AI art community.

The catch: "open" has limits

Two critical pieces of H3 remain proprietary. The 2K resolution upscaler -- the module that lets the model output 2K video -- is closed. So is H3-Context-IR, which translates prompts and reference material into a structured intermediate format that the model generates from. Without Context-IR, users running H3 locally must handle prompt structuring themselves using published guides.

This means local inference tops out at 768p resolution. The full 2K experience requires MiniMax's cloud API, which is priced at roughly one-third of competing services. According to Pandaily, the API runs at CNY 0.8 per second of video -- about $0.11 at current exchange rates, compared to Sora's roughly $0.30 per second.

The license adds another constraint: commercial use is only permitted for companies with under $20 million in annual revenue. Larger enterprises must negotiate directly with MiniMax.

The competitive landscape heats up

H3 launched on the same day ByteDance released Seedance 2.5, a closed model that generates 30-second clips with built-in audio. The timing was not coincidental -- both Chinese AI companies are racing to dominate a market that was, until recently, considered the exclusive domain of US firms.

Alibaba also released Qwen3.8-Max on the same day, making August 3 a concentrated splash of Chinese AI releases. The Information reported that MiniMax's H3 launch is part of a broader strategy to position itself as the "DeepSeek for multimodal AI" -- open-weight, aggressively priced, and targeting developers who want to build on video AI without lock-in.

MiniMax stock surged nearly 13 percent on the Hong Kong exchange following the announcement.

What this means for the video AI landscape

The idea that frontier video generation requires proprietary, API-only models has been the industry assumption since Sora's debut. H3 challenges that assumption directly. An open-weight model that ranks competitively on standardized benchmarks means developers can fine-tune on custom footage, adapt to specific visual styles, and run inference on their own hardware.

AMD ROCm support at launch (reported by blockchain.news) means the model runs on GPUs from both Nvidia and AMD -- a notable step toward hardware diversity in AI video inference. Two RTX 5090 cards are enough for 768p local generation.

The downside is real. Without Context-IR, local users face a significantly worse experience. The 2K gap means the best output is still cloud-only. And the $20M revenue cap on commercial use means most startups will need to negotiate or wait for the license to evolve.

But the trajectory is clear. Every closed model that gets topped by an open one loses a piece of its aura of inevitability. Sora was supposed to be the ceiling. H3 just proved the ceiling is lower than anyone thought.

Sources

• MiniMax H3 on HuggingFace
• The Decoder: China's MiniMax H3 is the first open model to top an AI video ranking
• Startup Fortune: MiniMax's H3 Video Model Undercuts Sora and Veo on Price and Openness
• SCMP: Video AI: MiniMax challenges ByteDance with low price, open weights
• Pandaily: MiniMax Releases Open-Source Full-Modal Model H3
• The Information: China's MiniMax Launches New Open-Source AI Video Model
• Artificial Analysis: Video Model Leaderboard