Two of the most capable AI models on the market in mid-2026 were built almost as opposites. OpenAI keeps the numbers that power its GPT-5.6 Sol locked inside its own servers. Moonshot AI, a Chinese startup, releases the numbers behind its Kimi K3 model for anyone to download and run. Both ranked near the top of independent tests that compare AI models.
Those numbers are called weights. During training, a model is shown enormous amounts of text, code and images, and its billions of weights are adjusted again and again until its answers improve. Picture each weight as a small dial whose setting holds a piece of what the model has learned. Together, those settings separate a model that only flags a bug from one that also explains the cause, the fix and why it works. Whoever controls the weights controls that difference in judgment.
OpenAI and Anthropic, two of the largest AI developers, keep theirs private, because publishing them would hand a rival years of training investment at no cost. Businesses reach these models through products like ChatGPT or Claude, or by connecting their own software to the company’s servers and paying for each use. Either way, the model never leaves the company’s control, so no outsider can download it, adjust it or run it on their own hardware. The industry calls this a closed weight, or proprietary, model.
DeepSeek, Alibaba and Moonshot AI have published, or promised to publish, the trained weights behind some of their models, a practice called open weight. Once the weights are public, any business can download the model, run it on its own computers, adjust it for a specific job and study how it behaves. It can also build products on the model without depending on the original company’s service or pricing. These developers, many of them based in China, were catching up with American labs while charging less.
Prices fall because more than one company can host a public model, even though a large one still demands serious computing hardware. No single company owns the right to serve it, so cloud providers compete to run the same model as cheaply and reliably as possible.
Washington noticed the growing reliance on cheaper open weight models built in China. By July 2026, officials had discussed restricting or banning them on national security grounds, without reaching a decision.
