Most people know ChatGPT, Claude, and Gemini. Fewer are familiar with Kimi, GLM, or DeepSeek. That second list matters more than it used to. Last week Moonshot AI, a Chinese AI company, released Kimi K3. It's an "open-weight" AI model that now performs close to the best "closed-weight" AI models built in the United States. The gap that long separated the two camps has narrowed considerably. But what exactly is the difference between "open-weight" and "closed-weight" AI?
A "closed-weight" model is like renting a house. Everything is ready on move-in day. The furniture is provided, the WiFi is set up, and someone else handles repairs. The trade-off is that the place cannot be reconfigured to better suit its occupant and the landlord sets all the terms.
On the other hand, an "open-weight" model is more like being handed a house for free. It can be gutted, rewired, and remodeled into whatever its owner wants. The catch is that the owner is responsible for all the money and effort that goes into upkeep.
Since the launch of ChatGPT, the house someone could rent was much nicer than the house someone could own for free. However, in recent months, the difference in the quality of the house that could be rented or owned has shrunk considerably. As a result, companies are increasingly reevaluating their options.
The Trade-Offs
The closed-weight "rental" model carries real advantages. The AI company handles the technical work in order to ultimately provide a simple set of AI tools almost anyone can use. Anthropic, for instance, now offers tools like the Claude Cowork desktop app, plugins for Microsoft Office, and an assistant inside Slack. The cost is control. Access can be restricted at any time for any lawful reason.
Since open-weight models can be "owned" directly on a computer or server, it can avert many of those restrictions. It can also be further personalized. In a recent example, Bridgewater Associates fine-tuned the Chinese open-weight model Qwen3-235B to better reflect the deep wisdom from its team. The result? The average task accuracy jumped to 84.7% (versus 78.2% for the leading frontier models) and the inference cost was 13.8x cheaper. With that said, there still was the time and monetary cost that went into fine-tuning the AI model.
Which Way Is the Wind Blowing
China is not the only country with open-weight models. The U.S. has some too. Google has Gemma. Microsoft has Phi. OpenAI has gpt-oss. Some of them are quite good too. However, those companies also have a proprietary, closed-weight AI model they sell too. As a result, it makes sense that these open-weight models would underperformed their closed-weight (revenue-generating) counterparts.
Nearly every month a new one tops the charts. In some ways the open-weight and closed-weight AI models have become commodities. What may matter more is how the AI gets applied. OpenAI, Anthropic, and Microsoft all recently launched consulting companies specifically focused on deploying AI. It suggests that it's good to have smart AI models, but smart AI deployments may ultimately make the biggest difference on the business results that follow.