Hosted providers have filtering on their chat interfaces (you can see how the response starts streaming in), doesn't mean the model itself has that behaviour.

Plus, your video is of Kimi K2.6.

Your video is showing Kimi 2.6, not 3? Once the weights are released, there might be providers that serve it without the censorship filter.

If this is open weight, would abliteration counter the refusals?

> It’s possible that the chat/ux model does know and has an unbiased opinion about china, but the filtering is on the front end/client side and so the user facing model has not been fine tuned natively

This has always been the case with Chinese models. Their web ui is a service provided from China, so they are required to censor it, but not the models themselves.

What does it say if you ask it in Chinese? I imagine English training data would be more critical of China than Chinese

There is no such thing as an unbiased opinion, opinions always contain some sort of value judgement. Besides, training data contains biases.

What is plausible is that they haven't made any attempts to explicitly steer certain opinions into a certain direction, and just let the model take over the bias of the training data, whatever that may be. Filtering in the front end is the easy, lazy way out to be legally compliant.