MiniMax H3 open-source: Chinese lab breaks the price of full-modal

In plain terms

MiniMax launches H3, a full-modal model (text, image, audio, video) that the editor claims is ranked #1 worldwide in video editing, at a third of the price of competitors. In other words: the Chinese laboratory that suffered a painful share lockup in June retaliates with a product that repositions its commercial argument - not frontier LLM, but open multi-modal orchestration.

Ce qui change

Pandaily press release of July 31, 2026, MiniMax product communication. The H3 model, published open-source, combines four modalities in a single backbone. The claim of leadership in video editing is highlighted, without any known third-party benchmark as of this date. The price - a third of that of the cited competitors - is the real market signal: MiniMax seeks to do for video what DeepSeek did for text, break the price by pure positioning choice.

Under the hood

The full-modal open-source raises two immediate technical questions: the scale of the weights (to be monitored on HuggingFace) and the effective inference. A full-modal model runs a multi-head router plus specialized heads - the GPU cost per second of video generated depends on the size of the heads, not just the backbone. The displayed price only makes sense with in-house inference; in self-hosting, the actual bill will depend on the hardware and cluster usage.

So what

Two readings. AI video market: the Latent Space report of July 24 placed FLUX 3 (Black Forest Labs) ahead of Seedance 2.0 and Gemini Omni; H3 inserts itself below in price, above in modalities covered. The market stops being binary (expensive closed / weak open). Open-model-economics thread: MiniMax bets on volume subsidy and hardware diversification, like Alibaba with Qwen. The closed pivot predicted by Lambert is not the only viable trajectory. To watch: real adoption among creative tools (Runway, Pika, CapCut), independent third-party video benchmarks, and transparency on weights and licensing.