Generative video has lost its status as a closed-code stronghold. MiniMax, the Shanghai-based Chinese lab, released H3, its most advanced multimodal model, and announced that its full weights will be available in the coming days under a community license. The decision is no philanthropic gesture: it is a strategic maneuver timed to coincide with ByteDance’s launch of Seedance 2.5 on the same day and a price cut by OpenAI. Three moves, one battle: who controls the next major layer of generative artificial intelligence. And this time, the battlefield is neither text nor image, but video—the format that consumes the most compute, the most data, and the most energy. The question hanging over the sector is whether open weights—the strategy that already democratized language models—can achieve the same disruptive effect in a domain dominated by proprietary giants.
MiniMax’s strategic pivot: from proprietary model to open ecosystem
H3 is not just another model. According to the benchmarking platform Artificial Analysis, it is currently the most powerful AI model in the world for video editing, a claim that places it ahead of ByteDance’s Seedance 2.0 in certain specific tasks, although it trails Google’s Gemini Omni Flash in text-to-video and Seedance 2.0 and Gemini Omni Flash in image-to-video. Nuances matter: H3 is not the absolute leader in every category, but it marks a turning point for a reason unrelated to benchmarks.
The decision to open the weights breaks with the company’s own trajectory. The previous models in the Hailuo family were proprietary; H3 is the first to open its architecture. MiniMax’s justification, laid out in its blog, is explicit: closed-code models have long dominated video generation, with slower iteration and a less open ecosystem than fields like large language models. It is a declaration of principles that follows the script of recent history: open weights in language models—from Llama to Qwen or DeepSeek—accelerated innovation, cut costs, and multiplied use cases. MiniMax wants to replicate that effect in video, the format where the barrier to entry was highest.
The chosen license, the MiniMax Community License, is designed to maximize adoption without abandoning the business: free non-commercial use and free commercial use for organizations with annual revenues below a certain threshold, with attribution requirements. In practice, MiniMax ensures that large corporations pay for the API, while startups, researchers, and independent creators can build on its technology at no cost. It is the same model that other Chinese companies have already validated, and that has proven to be an ecosystem-generating machine.
ByteDance responds the same day: Seedance 2.5 and the multimodal war
The timing is no coincidence. ByteDance unveiled Seedance 2.5 also on July 31, 2026, in what appears to be a direct response to MiniMax’s move. According to specifications published by TechNode, ByteDance’s model can generate a high-quality 30-second video clip in a single run, supports multi-turn extension for longer sequences, and allows users to provide up to 30 images, 10 videos, and 10 audio clips as reference material in a single input. These are impressive capabilities, but the key difference is strategic: Seedance 2.5 remains within ByteDance’s closed ecosystem.
The company led by Zhang Yiming has confirmed the adoption of Seedance 2.5 by a significant group of industrial and tech companies, according to Chinese outlet 36Kr: XCMG, the heavy machinery giant; XPeng Motors, the electric vehicle maker; Zhiyuan Robot, the robotics startup; and others including Lingchu Intelligence, Weifen Zhifei, Qiongche Intelligence, and Xspark AI. The list reveals ByteDance’s strategy: not just competing in entertainment or content creation, but positioning its video model as infrastructure for industrial, automotive, and robotics sectors. Generative video is no longer a marketing tool but a component of production systems.
The implementation of Seedance 2.5 on the Jimeng AI and Doubao Pro platforms confirms that ByteDance will integrate the model into its mass-market products, while Volcano Engine, its cloud services arm, will launch the API shortly. It is the complete ecosystem: model, platform, cloud, and industrial clients.
OpenAI cuts prices: Chinese pressure felt in San Francisco
On the other side of the Pacific, OpenAI has responded to competitive pressure with a classic tool: price. Sam Altman announced on X a drop in the developer cost for the lightweight model GPT-5.6 Luna. The reduction, which affects the entire GPT-5.6 line, is the clearest admission to date that the low-pricing strategy of Chinese labs is eroding OpenAI’s commercial advantage.
The move has multiple readings. On one hand, it is a defensive response: developers who once paid for OpenAI’s API now have Chinese alternatives of comparable quality at a fraction of the cost. On the other, it is a sign that the AI economy is entering a phase of forced deflation, where margin per token is compressed and differentiation must come from quality, integration, or exclusivity. The problem for OpenAI is that generative video—precisely the ground where MiniMax and ByteDance are advancing—is one of the segments where its advantage is least clear.
Hardware, sovereignty, and the new geopolitics of generative video
The piece connecting all these moves to the digital sovereignty agenda is MiniMax’s statement on H3’s design. The company cites compatibility with a wide range of hardware as a key consideration in the model’s development. In a context where access to NVIDIA’s latest-generation chips is restricted for Chinese companies, this line is more than a technical note: it is a declaration of infrastructural independence. A model that works well on diverse hardware—not just the most advanced accelerators—reduces dependence on a single supplier and expands the potential deployment market.
For public officials and European companies, the lesson is twofold. First, open weights in generative video opens the door to local, sovereign deployments, without relying on the APIs of American or Chinese giants. Second, the price war between OpenAI, ByteDance, and MiniMax is compressing the costs of a technology that until recently was prohibitive for most organizations. A high-quality video model with open weights, runnable on available hardware, is exactly the kind of tool that can democratize audiovisual production in mid-sized markets.
The final reflection points to the immediate future. If the history of large language models repeats itself, the opening of H3 will not be an isolated event, but the start of a cascade: other labs will be forced to open their video models to avoid missing the ecosystem train. The foreseeable consequence is an acceleration of innovation in editing, generation, and post-production tools, but also a fragmentation of the market where the competitive advantage will no longer lie in the model itself, but in integration, proprietary data, and compute capacity. Generative video has ceased to be a laboratory luxury and become strategic infrastructure. Whoever controls its open ecosystem will control a significant share of the next digital decade.