Because sometimes (a lot of the time in my experience) third-party providers and inference engines fail to implement the model correctly in ways that are sometimes very subtle and not obvious.
Deepinfra for example is not preserving thinking correctly for GLM5.1, even though they are for GLM5. This is one of the more obvious issues that crop up.
Yeah some of the providers on Openrouter correctly list what quantization they offer. And some refuse to say. OR should kick them off their platform if they want to be secretive.
Deepinfra for example is not preserving thinking correctly for GLM5.1, even though they are for GLM5. This is one of the more obvious issues that crop up.