These days, I have been connecting a small tool to a model and have changed four or five before and after. To be honest, there is no shortage of AI models nowadays. When you open any platform, the names can be listed on two pages. What's really troublesome is which one is more useful for the other thing? Some answers are indeed good, but calling once is not cheap. Some models have low prices, but they start to get stuck during peak hours. There are also some models that have very strong reviews and are really stuck in their own workflow. After running three times, they can get two wrong answers I just saw that the Model Marketplace for @ dgrid_ai was launched, and my first reaction was that I might have made another model shelf, right? Put up a bunch of model names, this thing is really not scarce now. But what DGrid wants to deal with this time is not just where to find the model, but how the model runs from being put on the shelves to being used in real life. The model provider can connect the model, and the inference node is responsible for providing computing power and services. Developers can call it according to their needs, and finally users can actually use it. Previously, most of these processes were locked in centralized platforms, where the platform decided who to recommend, who to give traffic to, and how to set prices. Even if a model made by a small team has good results, without platform resources, it may still be buried behind the list. DGrid aims to enable models and inference services to directly meet the real needs of developers, without relying solely on platform editors to click on recommendations in the background. I am more interested in the AI Arena here. It's not about asking everyone to vote based on the model name, nor is it about making another ranking of who is the world's number one. Users see answers from different models, but do not know who is behind them, so they choose based on their own real experience. This detail is quite important because once the brand name is revealed, it is easy for people to have preconceptions. Seeing the big factory model, I subconsciously feel that it is stronger. Seeing a model that I have not heard of before, I have already deducted points before reading the answer. Blind testing should at least remove this interference first. Users don't need to understand parameters, scores, and technical reports, they just need to answer a simple question about which of these two results I actually prefer to use. Next is PoQ, Proof of Quality. Decentralized AI used to prefer saying that the more nodes, the better, but having a large number of nodes only indicates that there are many machines, not necessarily reliable services. A node runs fast today and starts to drop tomorrow. The quotation is very cheap, but the delay is high and the results are unstable, which still cannot be used by developers. PoQ focuses on how to verify the quality of inference services and how to obtain more reasonable incentives for high-quality and efficient services. It's not that whoever comes first and has more machines will always receive rewards, but rather that the network should gradually align resources towards truly deliverable services. Putting these pieces together makes the logic more complete. Model Marketplace is responsible for discovering models and services, AI Arena sends back real user preferences, and PoQ processes inference quality and incentives. The model selected by the user may not necessarily be the one with the largest parameters, but may be the one that is more suitable and cost-effective for the current task. Developers don't have to bet on a certain brand every time, they can make choices based on real feedback and service quality. This is much more interesting than promoting our many models separately, because the number of models is no longer a problem. How to deliver suitable models to the right needs is what is currently lacking. And DGrid didn't start this closed loop right after setting up its official website. The official previously disclosed a six-month 23 million ARR, which at least indicates that the network has been exposed to real business and payment needs. Now launching the Model Marketplace is more like continuing to supplement the market layer on top of the original inference business, integrating model supply, computing power services, developer calls, and user feedback together. Of course, I wouldn't just say that a product image is working just because of it. The market ultimately depends on whether the model supply is abundant enough, whether the inference service can be stable in the long run, whether user feedback can really affect discovery and routing, and whether good nodes ultimately earn more. But I recognize the direction. The next step for decentralized AI is no longer just about who has more nodes and who has more promotion. The model is not valuable when it is put on the shelf, but it is valuable when it is discovered and called by real users, and then gains more demand because it is truly useful. What DGrid wants to fill in now is the middle section. As for whether it can ultimately become a truly open AI inference market, I will continue to observe and put it into practical use. After all, the quality of the model can be determined by running it a few times.