AI HOT (Curated Pool)Jul 31, 2026, 17:59 UTC+8

MiniMax launches open-source H3, a multimodal model that generates 2K video with native stereo audio

MiniMax H3 发布：开源全能多模态生成模型，支持 2K 原生立体声视频

MiniMax H3 unifies text, image, video, and audio understanding into one model. It generates up to 15-second videos at 2K resolution with native stereo sound. The company claims per-second pricing at 2K is under one-third of mainstream models, and 768p is under half the price of mainstream 720p. Model weights will be open-sourced in the coming days, subject to legal review. The post does not specify the open-source license or exact release date.

Why it matters: MiniMax dropped H3, a multimodal generation model that handles text, image, video, and audio in one model, outputting 15-second 2K video with native stereo sound. Pricing is aggressive — 2K per second at less than one-third of mainstream models — and weights are planned for op...

Read the original ↗Export Markdown

Take

This one's worth opening because MiniMax is swinging on three fronts at once: multimodal understanding, native stereo audio, and aggressive pricing. H3 takes images, audio, and video as inputs, then stitches them into a single 15-second 2K clip with stereo sound based on your text prompt. The company claims per-second pricing at 2K is under a third of mainstream models, and 768p is less than half the price of mainstream 720p.

I'd discount those numbers a bit. The post doesn't name the "mainstream models" it's comparing against—Runway Gen-4, Kling 2.0, Sora? No way to verify. The open-source promise is also vague: "weights in the coming days, subject to legal review," with no license specified and no exact date. If it lands under a non-commercial license, the open-source label loses a lot of weight.

Two things I'm actually watching: first, whether the cross-modal instruction following holds up in production. Referencing a Hitchcock camera move from one video, a character from an image, and vocals from an audio clip sounds impressive in a demo, but ad and e-commerce workflows will stress-test it fast. Second, hardware requirements. The post says hardware compatibility was a priority from early design, but there's no VRAM number. If this runs 2K inference on a 24GB consumer GPU, the pricing advantage becomes real. If it needs an H100 cluster, it's just another API with a blog post.

Source: AI HOT (Curated Pool) · minimax.io
Published: Jul 31, 2026, 17:59 UTC+8 · Jul 31, 2026