Summary of Key Points
In the past, the weights of open-source large models could be used freely for free, and companies made money through APIs and private deployments. However, as the performance of open-source models has caught up with that of closed-source models, model companies have begun to charge those enterprises that are making significant profits from using these models by adjusting license terms (for high-income MaaS services and specific use cases) and raising API prices. This approach allows them to maintain an open ecosystem (where ordinary developers can still use the models for free) while seizing high-value commercial opportunities and preventing others from reaping benefits without contributing.
1. License Restrictions: Targeting Enterprises Making Large Profits from Open-Source Models
The licenses for open-source models are no longer uniformly free. Instead, they impose barriers for high-income commercial scenarios (especially MaaS services):
- Kimi K3: Allows free download, fine-tuning, and commercial use. However, if a company provides MaaS services using Kimi K3 and its total revenue exceeds $20 million, it must sign an agreement with Moony Face. The focus is on the company's overall scale, not just the revenue generated from Kimi. For example, large cloud providers using Kimi for services may trigger these conditions, even if they are just starting out, while smaller startups with $19 million in revenue are exempt.
- Qwen3.8-Max: In addition to MaaS, using it for AI office assistants (such as programming or office tools, similar to Alibaba’s own Qoder/QwenWork) is prohibited, and the threshold is even higher ($50 million in total revenue). However, smaller models like Qwen3.8-27B remain completely open, reflecting a strategy of using flagship models to attract an ecosystem while keeping smaller models accessible.
- MiniMax: Initially, commercial use was prohibited, but later it was allowed for commercial products with annual revenue below $20 million (as long as the source is credited). The definition of commercial use remains somewhat ambiguous.
- Special Cases: MiniMax’s video model H3 is excluded from the US, EU, UK, and Korea due to copyright lawsuits (in the US) and regulatory uncertainties (in the EU and other regions), and companies in these areas need to apply for separate licenses.
These terms essentially mean that ordinary developers can use the models for internal tools and small applications without issues, but if they turn them into a large-scale business (especially by selling model services), they must share the profits with the original company.
2. Other Ways to Make Money Besides Licenses
Not all companies rely on license restrictions. Here are two other approaches:
- DeepSeek: The model weights are still free, but the prices of their API services have skyrocketed (by 350% during peak times). The reasoning is that users who deploy the models themselves need to handle the computational resources and stability, which is costly. DeepSeek has a cost advantage, so even with the price increase, users may still choose their services.
- Zhipu: Their new generation of models uses the MIT license (completely open), but they make money through private deployments (accounting for 73.7% of total revenue) and API services (26.3%; after raising prices by 83%, usage increased by 400%). The license reduces the barriers for developers, expanding the ecosystem, while high-value enterprise services generate direct revenue.
- MiniMax: They develop C-side AI applications (such as chat and video generation) and also provide the models to developers, managing third-party commercial use through licenses. This approach allows them to diversify their revenue sources.
3. The Background of These Changes
The main reason for these changes is that inference services are becoming increasingly valuable:
- Third-party inference platforms (such as Together and Silicon Flow) earn money by running open-source models.
- Scenarios like coding and agents have led to a significant increase in token consumption (a single task may now use ten times the amount of tokens previously).
- The company that deploys the model, performs the inference, and receives the revenue directly affects the model company’s commercial potential.
For example, if a large cloud provider uses Kimi’s open-source model to generate services and makes $1 billion without paying Moony Face, Moony Face will not be happy. This is similar to the situation in 2018 when companies like MongoDB and Elastic changed their licenses—cloud providers hosted their open-source products for free but made substantial profits, so they wanted to recoup their investment through licenses. AI models are facing the same issue now.
4. Can These Changes Really Generate Revenue? What’s the Current Situation?
We are still in the early stages, and the revenue has not fully materialized, but there are some developments:
- Moony Face has partnerships with Zhongruan International and DigitalOcean (the split ratio is not publicly known, but it is said to be up to 30%).
- They are in talks with Microsoft, Amazon, and Google about revenue sharing for K3 (aiming for 30% of related service revenue).
- The licenses mainly affect large platforms (such as Volcano Engine and Alibaba Cloud); ordinary developers and small companies are largely unaffected.
In essence, these changes create a negotiation framework: the weights can be obtained for free, but as the scale expands, companies need to sit down and discuss how to share the profits. While these changes have not yet become a significant source of revenue in financial reports, the direction is clear: open-source models are not a “free lunch,” and high-value commercial scenarios require payment.
5. The Essence: Balancing Open Ecosystem and Commercial Interests
These changes are similar to those in game engines like Unreal Engine, which are free for all developers to use. Once users start making money (for example, when game revenue exceeds a certain threshold), a share is taken. Model companies aim to:
- Attract more people to use their models through open licenses (to expand the ecosystem and increase their popularity).
- Charge those enterprises that are making large profits from the models (especially in MaaS and direct competitive scenarios) to ensure their own commercial returns.
The issue of how open-source models can generate revenue was previously unclear (mainly through APIs and deployments). Now, as the value of inference services increases, it has become more concrete: they want to keep the ecosystem thriving while preventing others from using their core assets for free.
In summary: Open-source models are no longer freely available for everyone. The approach is now “free for general use, but profits are shared for large-scale businesses.” Model companies have finally found a way to make money within an open ecosystem.