智东西（公众号：zhidxcom）
作者 | 江宇
编辑 | 漠影
硅谷AI公司，开始站在中国开源模型的肩膀上造模型了。
最近，出手的不仅有英伟达的技术团队，还有美国AI视频明星公司Pika。它们不约而同地选择了同一个中国视频模型——MiniMax H3。
英伟达相关技术团队基于MiniMax开源的视频模型H3，将低分辨率视频再生成至2K、4K，推出视频增强项目Open-H3-Regenerate-2K。几乎同一时间，Pika发布Pika H3 Flex，让模型以更少的推理步骤生成视频，并针对专业创作场景增强控制能力。
值得注意的是，英伟达等相关项目还涉及和MiniMax的授权。模型权重虽然已经公开，但依然有商业价值。
一家中国AI公司的模型，为什么会成为海外同行继续研发的起点？当越来越多企业选择基于开源模型创新，视频生成行业的技术竞争和商业分工又将发生怎样的变化？MiniMax H3的这轮生态扩张，或许提供了一个观察窗口。
一、英伟达、Pika接连下场，硅谷明星公司选中MiniMax
当前的AI行业里，海外开发者和公司正在尝试另一种路径：在现有的SOTA开源模型上继续创新。
今年8月，MiniMax正式开源H3。当时，全球AI模型评测机构Artificial Analysis公布的榜单中，其便在视频编辑任务上排名第一，达到当时的SOTA（业界最优）水平。 三个多星期后，该模型累计下载量已超过2400万次，产生超过300个公开衍生模型。H3作为基座模型，内容生成质量很高。
H3开放权重后，个人开发者和研究团队尝试让它在不同硬件上运行：ComfyUI背后的Comfy Org团队降低了模型运行时的内存占用，使其适配更多消费级显卡。Redis创始人antirez开发了h3.c，让H3能在苹果芯片的Mac上原生运行。FastVideo是加州大学圣迭戈分校Hao AI Lab发起的开源项目，团队与合作方通过蒸馏和稀疏化加快生成速度，也做了硬件适配。这些工作降低了运行门槛，为后续开发和微调提供了工具。
商业公司也在基于H3继续研发。英伟达相关团队尝试提高高分辨率视频再生成的效率，Pika通过后训练减少推理步骤、增强创作控制能力。
对他们来说，H3 提供了一个有多模态生成能力的底座。其他公司可以把研发投入放在后训练、生成速度、高分辨率和专业创作需求上，针对特定创作场景继续优化。
这是开源生态中的一种独有的产业分工：有人负责提供通用模型能力，有人负责训练专业模型，有人负责提升计算效率，还有人负责构建面向用户的产品。当其他公司开始围绕一款模型持续投入研发资源，这款模型便有机会从单一产品成为整个技术生态的组成部分。
二、一个开源模型如何做商业化？
长期以来，行业里对开源模型的疑问之一是开放权重有利于扩大技术影响力、吸引开发者，但模型厂商能从开源模型上获得可持续的商业回报吗？
H3选择了一条兼顾模型开放与商业化收入的路径。
目前，MiniMax H3采用专门的社区许可协议，并非Apache 2.0等传统宽松开源许可证。根据其公开许可协议，在许可适用地区内，符合条件的用户可以免费使用H3开展研究和商业活动。
但对于规模较大的商业使用，协议保留了进一步授权的要求。例如，若使用者的商业产品和服务年收入超过2000万美元，则需要事先获得MiniMax的书面授权。
此外，美国、欧盟、英国和韩国不在普通社区许可的适用范围内，相关地区的主体需要另行联系MiniMax获得授权。商业产品使用H3时，也需要按照许可要求明确显示MiniMax H3。
这意味着，H3可以通过相对低门槛的开放吸引开发者参与，同时为特定地区、大型商业客户和定制化需求保留独立授权空间。
此次英伟达相关项目基于H3开展的合作，同样是授权许可模式。
那为什么大公司愿意为已经公开权重的模型建立商业授权关系？
他们需要的往往不只是模型权重本身。除了获得明确的商业使用权限，企业还可能需要定制开发、技术支持、部署适配，以及长期合作的确定性。H3的开源为不同规模、不同需求的企业提供了定制调整、参与生态的方式。
从商业模式来看，MiniMax正在围绕H3探索多条相互关联的路径。
一方面，MiniMax可以通过官方API向开发者和企业提供模型服务，并通过MiniMax Design等产品触达终端创作者。
另一方面，针对具有特殊需求的大型企业，MiniMax还可以通过商业授权、定制模型和技术合作，提供更有针对性的服务。
这几条路径并非彼此割裂。开发者围绕H3探索更多应用场景，有机会带来新的企业需求；企业在实际业务中遇到的技术问题，又可以推动模型优化和产品迭代。随着应用范围扩大，模型服务、商业授权和定制合作也可能相互促进。
对MiniMax而言，H3正探索让模型在开放环境中被更多人使用，同时保留面向大型企业的商业合作空间。
这条路径的重要之处，是它尝试改变开源与商业化之间的关系：模型可以被广泛复用，商业价值则可能来自围绕模型建立的服务、授权和长期合作。
三、AI视频的下一场竞争中，谁来定义底座？
很长一段时间里，AI视频模型的进步有着一套清晰的衡量标准：画质越来越高、时长越来越长、生成速度越来越快。
这些指标仍然重要，但它们越来越难以回答另一个问题：一款视频模型，究竟能够创造多大的产业价值？
MiniMax H3开源后的变化，提供了一个新的观察角度。对整个AI视频行业而言，模型厂商之间的竞争，可能不再只看谁的模型更强，还要看谁的模型能够成为其他公司继续创新的起点。
如果越来越多公司不再从零训练基础模型，那么，整个行业的竞争格局也可能随之改变。
相比大语言模型相对成熟的开源生态，视频生成领域因为有大量技术门槛、使用门槛，即使开源，也不意味着普通开发者就能轻松使用，更不意味着围绕它的开发者生态能够自然形成。
可以明确的是，开源底座的优势在于生态，开发者积累。一旦大量工具、教程、推理框架和专业模型围绕某一基座建立，后续开发者便有更多理由优先采用它。
目前，H3拥有大量衍生模型，这虽然并不能直接证明其已成为行业标准。但英伟达相关团队和Pika等专业技术主体的加入，确实提高了它成为重要开源技术底座的可能性。 H3正在从一个开源模型，进一步成为下游企业开展技术创新的基础。
当越来越多公司的产品、技术和业务建立在H3之上，MiniMax也就更有机会将开源带来的技术优势，转化为对产业生态的长期影响力。
对MiniMax而言，H3已经迈出了吸引全球开发者参与的重要一步。但能否进一步成为被行业长期采用的基础模型，仍取决于其技术迭代、授权机制和商业合作能否持续经受检验。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Intellectual (WeChat Official Account:zhidxcom)
Author: Zheng Zheng, Jiang Woo
Edit |
Silicon Valley AI began to build models on the shoulders of the China open source model.
Recently, not only did it involve the technical team of Inverda, but also the American AI Video Star company Pika. They chose the same Chinese video model, Mini Max H3.
On the basis of Mini Max open-source video model H3, the technology team in Weida regenerated the low resolution video into 2K, 4K and launched the Video Enhancement Project Open-H3-Regenerate-2K. At almost the same time, PikapublishPika H3 Flex allowed the model to generate video in less reasoned steps and to increase control over the professional creation scene.
It is worth noting that related projects such as In Weida also involve the authorization of Mini Max. Model weights, though publicly available, still have commercial value.
Why is the model of a Chinese AI company the starting point for continued research by overseas counterparts? How will technological competition and the division of business in the video-generation industry change when more and more enterprises choose to innovate based on open-source models? This round of ecological expansion of Mini Max H3 may provide an observation window.
One, Young Weida, Pika, next. Silicon Valley Stars.
In the current AI industry, overseas developers and companies are experimenting with another path: continuing to innovate on the existing SOTA open source model.
In August this year, Mini Max officially opened source H3. At that time, the Global AI Model Evaluation Body, Artificial Analysis, ranked first on video editing missions, reaching the then SOTA (best industry) level. More than three weeks later, the model has been downloaded more than 24 million times, producing more than 300 open derivative models. H3 is a base model with high quality content generation.
After the H3 open weights, individual developers and research teams tried to make it work on different hardware: Comfy Org team behind ComfyUI reduced the memory occupancy of the model at the time it was running to make it more consumer-grade GPU. The founder of the Redis Antirez has developed h3.c to allow H3 to operate on the Mac of the apple chip. FastVideo is an open-source project initiated by Hao AI Lab, University of California, and the team and the collaborator have accelerated their generation through distillation and thinning and have made hardware adaptation. These efforts lowered the operational threshold and provided tools for subsequent development and fine-tuning.
Commercial companies are also continuing research and development based on H3. In an attempt to improve the efficiency of the regeneration of high resolution videos, the team in Weida, Pika, through training, reduced reasoning steps and enhanced creative control.
For them, H3 provides a base of multi-modular generation. Other companies can continue to optimize their research and development inputs in terms of post-training, production speed, high resolution and professional creativity needs.
This is an industry-specific division of labour in open-source ecosystems: some are responsible for providing generic modelling capabilities, others for training professional models, others for improving computing efficiency and others for building user-oriented products. As other companies begin to invest their resources on a continuous basis around a model, the model has the opportunity to become an integral part of the overall technological ecology from a single product.
How can an open source model be commercialized?
One of the long-standing questions about open source models in the industry is that open weights are conducive to increasing technological impact and attracting developers, but can model manufacturers obtain sustainable commercial returns from open source models?
H3 has chosen a path that balances model opening with commercial revenue.
Currently, Mini Max H3 uses a specific community permit agreement, not traditional liberal open source licences such as Apache 2.0. Under its public licence agreement, eligible users can use H3 for research and business activities free of charge in the area to which the licence applies.
For larger commercial uses, however, the agreement retains the requirement for further authorization. For example, prior written authorization from Mini Max is required if the user has an annual income of over $20 million for commercial products and services.
In addition, the United States, the European Union, the United Kingdom and the Republic of Korea are not covered by ordinary community permits, and the subject of the relevant area needs to be authorized to contact Mini Max separately. When using H3 for commercial products, MiniMax H3 is also required to be clearly shown in accordance with licensing requirements.
This means that H3 can engage developers through a relatively low threshold of openness, while retaining independent authorization space for specific areas, large commercial clients and customized needs.
This British Wida project is based on H3 cooperation, which is also a licensing model.
Then why would a large company be willing to establish a business authorization for a model that already has public weight?
They often need more than just model weights. In addition to obtaining a clear commercial use of permission, enterprises may need customized development, technical support, suitable deployment and certainty of long-term cooperation. The open source of H3 provides customized adaptation and eco-participation for enterprises of different sizes and needs.
In terms of business models, Mini Max is exploring multiple interrelated paths around H3.
On the one hand, MiniMax can provide modelling services to developers and businesses through official API, and reach end creators through products such as MiniMax Design.
On the other hand, MiniMax can provide more targeted services through business authorization, customized models and technical cooperation for large enterprises with special needs.
These paths are not fragmented. Developers explore more applications around H3 and have the opportunity to generate new business demands; technical problems encountered by enterprises in their actual operations can also contribute to model optimization and product overlay. Modelling services, business authorization and customization collaboration may also be mutually reinforcing as applications expand.
For MiniMax, H3 is exploring ways to increase the use of models in an open environment while retaining commercial cooperation space for large enterprises.
The importance of this path is that it attempts to change the relationship between open sources and commercialization: Models can be widely reused, while commercial values may come from services built around models, authorizations and long-term cooperation.
iii. Who will define the bottom seat in the next competition for AI video?
For a long time, the progress of the AI video model has been measured with a clear set of measures: it has become more qualitative, longer and faster to produce.
These indicators are still important, but they are becoming more and more difficult to answer: how much industrial value can a video model create?
The change after Mini Max H3 opens provides a new perspective. For the AI video industry as a whole, competition among model manufacturers may no longer depend solely on whose models are stronger, but also on who can become the starting point for other companies to continue their innovation.
If more and more companies no longer operate from the zero training base model, then the whole industry ' s competitive pattern may also change.
Compared to the relatively mature open-source ecology of the large-language model, the area of video generation does not mean that ordinary developers can easily use it because of the large technological thresholds, the use thresholds, and even the open-sources, not to mention that the developments surrounding it are naturally occurring.
It is clear that the advantage of the open source base lies in ecology and the accumulation of developers. Once a large number of tools, curricula, reasoning frameworks and professional models are built around a base, the successors have more reason to prioritize it.
At present, H3 has a large number of derivative models, although this does not directly prove that it has become an industry standard. However, the participation of the relevant teams in Weida and specialized technical subjects such as Pika has indeed increased its chances of becoming an important open-source technology base. From an open source model, H3 is further becoming the basis for technological innovation in downstream enterprises.
When more and more companies build their products, technologies and operations on H3, MiniMax will have a better chance of transforming the technological advantages of open sources into long-term impacts on industrial ecology.
For Mini Max, H3 has taken an important step towards engaging global developers. However, further development as a foundational model that has long been adopted by the industry depends on the sustainability of its technological overlap, empowerment mechanisms and commercial cooperation.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.