It feels like nitpicking, but there’s a relevant difference here. If you can only access the weights (and not the training data and the learning algorithms used), it’s hard to get (say) a Chinese model to freely talk about Tiananmen Square Massacre. (Just tried it locally on Deepseek R1; it’s doable, but takes some prompt hacking.) The models may be freely available, but there’s bias and censorship baked into them.
It feels like nitpicking, but there’s a relevant difference here. If you can only access the weights (and not the training data and the learning algorithms used), it’s hard to get (say) a Chinese model to freely talk about Tiananmen Square Massacre. (Just tried it locally on Deepseek R1; it’s doable, but takes some prompt hacking.) The models may be freely available, but there’s bias and censorship baked into them.