A Chinese AI startup just made it cheaper to generate professional-quality video than ordering lunch in most major cities. MiniMax Group released its H3 multimodal video generation model on July 31, capable of producing 2K resolution videos with stereo sound at what the company claims is less than one-third the cost per second of mainstream competitors.

What MiniMax actually built

The H3 model generates videos up to 15 seconds long at native 2K resolution. The model accepts text, images, videos, and audio simultaneously, meaning you can feed it a brand logo, a script, a reference clip, and a soundtrack and get back something coherent.

The practical applications lean heavily commercial: advertising, branding, e-commerce product videos, and gaming assets. MiniMax is positioning H3 not as a toy for social media creators but as a production tool for businesses that currently spend thousands on video content.

On pricing, the company claims H3 produces 2K video at under one-third the cost per second of mainstream models. For lower-resolution 768p output, the cost drops to under half of what competitors charge for 720p equivalents.

MiniMax plans to release the model weights as open-source in early August. That means anyone with sufficient hardware can run H3 locally, fine-tune it for specific use cases, and build products on top of it without paying MiniMax a dime.

The company behind the model

MiniMax was founded in December 2021 in Shanghai. It went public on the Hong Kong Stock Exchange in January 2026, raising approximately $619 million in its IPO. The company counts Alibaba and Tencent among its backers.

The H3 release follows MiniMax’s M3 model launch in June 2026, establishing a rapid cadence of releases designed to keep pressure on dominant players like ByteDance and Kuaishou. Both of those companies operate primarily closed-source video generation systems, making MiniMax’s open-source commitment a deliberate strategic differentiation.

Why crypto investors should care

A substantial portion of the AI-crypto thesis rests on the idea that AI compute and model access will remain expensive enough to justify decentralized marketplaces, token-incentivized GPU networks, and premium-priced inference protocols. Every time a company like MiniMax drops a state-of-the-art model at dramatically lower prices, and then open-sources it, that thesis gets harder to defend.

If you can run a 2K video generation model locally for free using open weights, the value proposition of paying tokens to access similar capabilities through a decentralized network becomes questionable. Decentralized compute networks still offer censorship resistance, permissionless access, and geographic distribution that centralized providers can’t match. But the premium those features can command shrinks when the underlying AI capabilities become cheap or free.

Investors watching the AI-crypto space should track two metrics closely. First, the gap between centralized AI pricing and decentralized alternatives. Second, the pace of open-source model releases from labs like MiniMax, Meta, and others. Each open-source release effectively commoditizes the previous generation’s capabilities.