> What does the ability to modify a model have to do with lock-in?

For example, Chinese models can spread the CCP propaganda and censor the results. You can't change it and may not even notice.

> Models aren’t software.

Models are software. They run on computers, require development, can have bugs etc. They can be downloaded or provided as SaaS.

> Models and inference are completely fungible.

Until the next version of DeepSeek becomes closed and you will only have the access to the last good open-weight model.

Sorry, are you actually paying any attention to the discussion? You’re spouting a bunch of buzzword non sequiturs. You started off by talking about how open weights aren’t open source when nobody said anything about open source. Now you’re talking about propaganda in response to a question about lock-in.

Lock-in means you can’t switch to a different model or a different provider because the cost (in time or money) of switching is high. Censorship and propaganda have nothing to do with it. You can switch to DeepSeek in an instant. You can switch away from it in an instant. The switching cost is zero.

And no, models are data files. The inference engines are software. Models run on many engines and most engines support many models. Another example of how there’s no lock-in: you can switch both models and engines at no cost. If you’re using an inference provider, they probably offer many models you can switch between, and you can switch to a different provider with no effort.

If the next version of DeepSeek becomes closed then you can switch to GLM or Kimi or MiniMax or whatever else in two seconds, or keep on using the last open version of DeepSeek. Zero lock-in.