Frontier, open weight, open source, Chinese, local. They don't measure along one scale.
You hear frontier, open weight, open source. What do all these terms mean? They don't measure along one scale. They describe different aspects of a model, and they matter when you're choosing the right AI solution for you.
Frontier means the best models right now, usually made by the frontier labs. That always includes Anthropic and OpenAI. It usually includes Google, Grok and sometimes Meta. It depends on who you ask. They have a couple of things in common. They're all closed models. You rent access, through a subscription or an API. You never get direct access to the model.
Open weight means the model file is public. Download it, run it, change it. You don't get the training data or the recipe. Real open source gives you all of it, but it's rare. When most people say open source about AI models, they really mean open weight.
When people say Chinese models, they mean models from Chinese labs. Most are open weight and near the frontier, typically within six months, and they're available at a fraction of the cost. That's their strategy.
Using a Chinese model means one of two things. You're using their app or API directly from the lab, and your data goes to their servers, under their rules and their law. Or you're using the open weights, run by somebody else. The model file can sit in another location, subject to different rules and laws. What doesn't change is how the model was trained and how it operates.
U.S. hosts often run these models, and you can access them from reputable U.S. providers subject to U.S. law. Many are also available under no-retention, no-training agreements, where your data is neither kept nor used to train the model. That gives you good data security while using open weight models.
Local models are models you run on your own machine. Nothing leaves your machine or your network, depending on how it's configured. You're limited by memory and processing power, so people typically use smaller open weight models like Qwen, Gemma or Llama. I use Qwen for processing here on my Mac mini. It's great for production and does a lot of work at no incremental token cost. It can be a wonderful solution for the right problem.
You probably already have a small model on some of your devices. Most modern iPhones and Macs come with Apple Intelligence, a small model running on the device that can be supplemented by a cloud model from Apple. Google does the same with Gemini Nano, and other providers follow the same approach. That gives you basic model access directly on your device.
Picking the right model means picking the right set of trade-offs: frontier performance, control of your data, keeping data domestic, low cost and overall access. Hopefully this helps you decide which model is best for your use case.