From Nvidia to Pika, Why Silicon Valley Companies Build on MiniMax H3

A Chinese AI company used Nvidia's GPUs and referenced Silicon Valley's technical approach to train its own large video model—few people would find that surprising.

But what if it were the other way around?

Nvidia, the world's leading GPU giant, and its research team made the rare choice to adopt the open-source multimodal model of a Chinese company, investing manpower and time to develop a video generation model. Now that is interesting.

From startups to Nvidia, why was H3 chosen?

In August this year, MiniMax released the weights of its multimodal model H3. This model can uniformly understand multiple types of material such as text, images, audio, and video, while balancing generation quality, speed, and cost. From a technical ecosystem perspective, opening the model weights allows external teams to directly participate in post-training, distillation, and inference optimization, expanding the ways they can take part in model development. Within one month of open-sourcing, H3 was downloaded more than 24 million times, with over 300 public derivative models developed based on it.

The first to be attracted were video generation companies. Most of their customers use video generation models to make advertising and marketing materials, and care about generation speed and the consistency of products and characters. No matter how dazzling a technical demo is, when it reaches the commercial production floor, customers tend to care about more practical questions: Will the text on product packaging become distorted? Can the same character remain unchanged when the camera angle changes? To modify one segment of an ad, does the entire video need to be regenerated? If thousands of short videos are to be produced in batches, can the time and cost be borne? These questions determine whether AI video is merely an awe-inspiring technology or a business that can truly operate at scale.

Competition in the industry is fierce, and these video generation companies are motivated to adjust models more specifically to customer needs.

Generative media infrastructure company fal started from these needs, training H3 Max on the basis of MiniMax H3 and optimizing the inference system around the model, with the goal of faster generation. Video generation company Pika wanted to better meet the needs of professional creators on its platform, and on October 29 it released H3 Flex, post-trained on the basis of H3, which compresses the generation steps.

Somewhat surprisingly, unlike these companies in vertical industries, Nvidia is developing a video generation model based on H3. In late October, the relevant Nvidia research team released Open-H3-Regenerate-2K based on H3, first using H3 to generate a low-resolution draft, then regenerating a high-resolution video through an additionally trained refinement module and three-step denoising. The project retains H3's original weights.

Nvidia rarely invests so many resources in multimodal models, but it has carried out systematic development on H3.

Previously, they launched the StreamRender-H3 streaming video rendering solution and a fast video decoder, exploring interactive generation and decoding acceleration respectively; the inference optimization project, meanwhile, improves the generation process while keeping the H3 model weights, prompts, and generation configuration unchanged.

Video generation companies and global giants like Nvidia have different application scenarios for models, but their standards for choosing a base model are both high: it must be able to handle complex multimodal inputs and video generation tasks, and also open its weights, allowing researchers to run the model locally, add refinement modules, or modify the inference process. Even if a model is fully open, if its generation quality is not good enough, its architecture is difficult to extend, or it cannot adapt to new training tasks, it will be hard for it to become the foundation for further R&D.

There are not many multimodal models that meet this standard, and H3's performance and extensibility have led it to be chosen repeatedly. In September this year, H3 ranked first on the Arena image-to-video leaderboard with 1,495 points, having received more than 57,000 evaluation votes in total.

English text is being prepared.

The multimodal model field may be the next CUDA-like opportunity: doing inference acceleration and generative applications around H3, bringing more video models and developers into the NVIDIA technology stack. What Nvidia is optimistic about is not only the video generation market, but also the ecosystem and computing power demand behind it.

In the open-source ecosystem, partners are also "nodes" for entering larger markets

When leading companies integrate video models into production processes, the requirements they face differ, and the same company will also raise new demands as its business changes. If model vendors satisfy these requirements one by one, while also continuously maintaining and updating, costs will keep increasing.

Research by video generation platform Runway found that some enterprise customers want to directly own the model weights, deploy the model in their own environment, and then fine-tune it with their own data. Interviews with enterprise AI leaders by venture capital firm a16z also mentioned that control and customization for specific businesses are important reasons why enterprises choose open models. It also noted in its report that enterprises are increasingly valuing open models, not only because of price, but also because they can be fine-tuned with their own data.

Open-weight models provide a new solution. Open weights give leading companies the conditions to put models into their own production systems. They can adjust models around existing customer needs, and then deliver capabilities through their own platforms, products, and services.

This points to a vast market far beyond what a single company can serve. Each of these partners connects customers, platforms, and workflows. For model and product vendors such as MiniMax, partners can become nodes for entering more industries. As H3 enters more customers and scenarios through partners' products, MiniMax can also continue to expand its own industry applications, enabling more technical teams familiar with vertical scenarios to participate, carry out targeted optimization based on H3's existing capabilities, and jointly promote the implementation of video generation technology in different fields.

This is also a commercialization approach unique to the open-source ecosystem. MiniMax can use these important companies as "nodes" to expand revenue sources through licensing, deployment support, or joint development.

This is a commercial path that has already been validated in multiple industries. Horizon Robotics, a smart driving chip and algorithm company, opens its BPU chip IP and perception algorithm licenses, and also provides chips, toolchains, solutions, and joint development. Automakers and partners can build differentiated products on this basis. This not only expands the company's "circle of friends" 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 the applicable territory, allowing users to use and modify the model and develop derivative models under the terms; if the annual revenue from commercial products or services using H3 exceeds US$20 million, prior written authorization from MiniMax must be obtained.

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

As video generation evolves from an occasional creative tool into a new link in enterprise production chains, directly related industries such as advertising, e-commerce, film and television, gaming, and education, as well as broader fields, will each spawn new tools and services. Open weights make H3 a foundation jointly transformed by these teams, and also give MiniMax the opportunity to enter more industries along partners' products.

This echoes the mission MiniMax has always stated, "Intelligence with Everyone": the development of AI is not just a few model companies continuously releasing stronger versions; more developers and more companies will also be nodes in this wave of technological innovation. Everyone is a user of new technology and also a driver of the wave.

For a commercial company, what matters more is that the market may reassess the commercial value of an open-source model: in addition to API revenue, when a model can become the foundation for enterprise customization, deployment, and product development, enter workflows in more industries, and drive revenue from licensing, technical services, and joint development, its commercial potential may be greater than imagined.

This article comes from the WeChat public account "AI Frontline" (ID: ai-front), author: Dongmei, published by 36Kr with authorization.