NVIDIA Builds Video Models on MiniMax's Open-Source H3

It would hardly surprise anyone if a Chinese AI company, using NVIDIA GPUs and following Silicon Valley’s technical playbook, trained its own large video model.

But what if the roles were reversed?

NVIDIA’s research team—a global leader in GPU technology—rarely chooses to invest significant manpower and time into developing a video generation model based on an open-source multimodal model from a Chinese company. This is intriguing.

From Startup to NVIDIA: Why Was H3 Chosen?

In August of this year, MiniMax released the weights for its multimodal model, H3. This model can uniformly understand diverse inputs such as text, images, audio, and video, while balancing generation quality, speed, and cost. From a technical ecosystem perspective, releasing model weights allows external teams to directly participate in post-training, distillation, and inference optimization, expanding the ways they can contribute to model development. Within one month of going open source, H3 surpassed 24 million downloads, with over 300 public derivative models built upon it.

The first to be attracted were video generation companies. Their clients primarily use video generation models for advertising and marketing materials, prioritizing generation speed and consistency across products and characters. No matter how dazzling the technical demos are, in commercial production environments, clients care about more practical issues: Will the text on product packaging distort? Can the same character maintain consistency when switching camera angles? Does modifying a segment of an ad require regenerating the entire video? If thousands of short videos need to be produced in bulk, are the time and costs sustainable? These questions determine whether AI video is merely a stunning technology or a business that can truly scale.

Amid intense industry competition, these video generation companies have strong incentives to tailor their models specifically to customer needs.

Fal, a generative media infrastructure company, started from these requirements, training H3 Max based on MiniMax H3 and optimizing the inference system around the model to achieve faster generation speeds. Pika, a video generation company aiming to better serve professional creators on its platform, released H3 Flex on October 29. Built via post-training on H3, H3 Flex compresses the number of generation steps.

Surprisingly, unlike these vertical industry players, NVIDIA is also developing a video generation model based on H3. In late October, NVIDIA’s relevant research team released Open-H3-Regenerate-2K based on H3. The approach first uses H3 to generate low-resolution drafts, then regenerates high-resolution videos through an additionally trained refinement module and three-step denoising. The project retains H3’s original weights.

NVIDIA rarely commits such extensive resources to multimodal models, yet it has undertaken systematic development around H3.

Previously, they introduced StreamRender-H3, a streaming video rendering solution, and a fast video decoder, exploring interactive generation and decoding acceleration respectively. Their inference optimization projects improved the generation pipeline while keeping H3’s model weights, prompts, and generation configurations unchanged.

Video generation companies and global giants like NVIDIA differ in their application scenarios for models, but their criteria for selecting base models are equally rigorous: the model must handle complex multimodal inputs and video generation tasks, offer open weights, and allow researchers to run the model locally, add refinement modules, or modify inference pipelines. Even if a model is fully open, it struggles to serve as a foundation for further development if its generation quality is insufficient, its architecture is difficult to scale, or it cannot adapt to new training tasks.

Few multimodal models meet these standards. H3’s performance and extensibility have led to its repeated selection. In September of this year, H3 ranked first on Arena’s image-to-video leaderboard with a score of 1,495, accumulating over 57,000 evaluation votes.

However, NVIDIA’s goals may extend beyond this. It excels at and dares to invest early in new technologies to cultivate ecosystems. One example is CUDA: by tilting resources and leveraging business strategies, NVIDIA enabled developers to build applications around its GPUs, forming a software ecosystem with high barriers to entry.

The multimodal model space might represent the next "CUDA-like" opportunity: building inference acceleration and generative applications around H3 to bring more video models and developers into the NVIDIA tech stack. NVIDIA sees potential not just in the video generation market, but also in the underlying ecosystem and compute demand.

In the Open Source Ecosystem, Partners Are Also "Nodes" for Entering Larger Markets

When leading enterprises integrate video models into their production workflows, they face varying requirements, and even the same enterprise develops new needs as its business evolves. If model vendors attempt to satisfy each requirement individually while maintaining continuous updates, costs will escalate indefinitely.

Research by Runway, a video generation platform, found that some enterprise clients prefer to own model weights directly, deploying them in their own environments and fine-tuning them with proprietary data. Interviews conducted by venture capital firm a16z with corporate AI leaders also highlighted that control and customization for specific business needs are key reasons enterprises choose open models. Their report noted that enterprises increasingly value open models not only for pricing but also because they can be fine-tuned with internal data.

Models with open weights offer a new solution. Open weights enable leading enterprises to integrate models into their own production systems. They can adjust the model around existing customer needs and deliver capabilities through their platforms, products, and services.

This points to a vast market far exceeding what any single company can serve. These partners each connect to customers, platforms, and workflows. For model and product vendors like MiniMax, partners can become nodes for entering more industries. As H3 reaches more customers and scenarios through partner products, MiniMax can continuously expand its industry applications, inviting more technical teams familiar with vertical scenarios to participate. By performing targeted optimizations based on H3’s existing capabilities, they jointly drive the implementation of video generation technology across different fields.

This represents a unique commercialization path within the open source ecosystem. MiniMax can leverage these key companies acting as "nodes" to expand revenue sources through licensing, deployment support, or joint development.

This is a proven commercial path across multiple industries. Horizon Robotics, an intelligent driving chip and algorithm company, opens up BPU chip IP and perception algorithm licenses, while also providing chips, toolchains, solutions, and joint development opportunities. Automakers and partners can use these to build differentiated products. This not only expands the company’s network and influence but also directly supports its commercialization.

According to the MiniMax H3 Community License Agreement, the agreement grants a royalty-free limited license within applicable regions, allowing users to use, modify the model, and develop derivative models according to the terms. However, if the annual revenue of commercial products or services using H3 exceeds $20 million, prior written authorization from MiniMax is required.

As the ecosystem becomes more active, the commercialization of MiniMax H3 is accelerating rapidly.

As video generation transitions from an occasional creative tool to a new link in the enterprise production chain, directly related industries such as advertising, e-commerce, film and television, gaming, and education—along with broader sectors—will each spawn new tools and services. Open weights enable H3 to serve as a common foundation for these teams to customize and improve, while also giving MiniMax the opportunity to enter more industries through its partners' products.

This echoes MiniMax's longstanding mission: "Intelligence with Everyone." The development of AI is not just about a few model companies continuously releasing stronger versions; more developers and more companies will be nodes in this wave of technological innovation. Everyone is both a user of new technologies and a driver of this momentum.

For a commercial company, what matters more is that the market may reassess the business value of open-source models. Beyond API revenue, when a model can serve as a foundation for enterprise customization, deployment, and product development—entering workflows across more industries and driving revenue from licensing, technical services, and joint development—its commercial potential may be greater than imagined.